{
  "community": [],
  "cross_family_note": "Each board is one result-sealed bundle, shown with its interval and its provenance. Benchmark families share no denominators.",
  "gateway": {
    "bundle_count": 4,
    "bundles": [
      {
        "arms": [
          {
            "arm_id": "direct-openai",
            "baseline": true,
            "conditions": {
              "cold": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.00147,
                    "p95": 0.0017947499999999997
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cold_end_to_end_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.5989910415373743,
                    "p95": 1.1089955310337243
                  },
                  "cold_end_to_end_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.5984453125856817,
                    "p95": 1.0626401207176968
                  },
                  "cold_end_to_end_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.3095027711242437,
                    "p95": 2.668665048014372
                  },
                  "cold_end_to_end_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.1131625415291637,
                    "p95": 3.9385627543553703
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 42.5,
                    "p95": 45.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.00147,
                    "p95": 0.0017947499999999997
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.00147,
                    "p95": 0.0017947499999999997
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 42.0,
                    "p95": 52.749999999999986
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.9197582711931318,
                    "p95": 3.7222186022903765
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.36092168744653463,
                    "p95": 0.8328393687726925
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.3607836253941059,
                    "p95": 0.7849101584870365
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.1458857920952141,
                    "p95": 2.405911793909035
                  },
                  "setup_dns_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.012593270977959037,
                    "p95": 0.1725526083493604
                  },
                  "setup_tcp_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.07028554216958582,
                    "p95": 0.218238822882995
                  },
                  "setup_tls_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.07810797891579568,
                    "p95": 0.2162797652184963
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 51.18058642803702,
                    "p95": 69.7586432758243
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 84.5,
                    "p95": 95.64999999999999
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              },
              "warm": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0030425,
                    "p95": 0.00336
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 42.0,
                    "p95": 45.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0014875,
                    "p95": 0.00176775
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0030425,
                    "p95": 0.00336
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 42.5,
                    "p95": 52.0
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.9094183123670518,
                    "p95": 3.97886997279711
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.3512618546374142,
                    "p95": 0.6831760892877352
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.35081070801243186,
                    "p95": 0.6830214519985016
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.0220603130292147,
                    "p95": 2.3214588018367053
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 47.920696368634346,
                    "p95": 70.47526051298266
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 85.0,
                    "p95": 93.55
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              }
            },
            "role": "direct"
          },
          {
            "arm_id": "cloudflare-openai",
            "baseline": false,
            "conditions": {
              "cold": {
                "availability": {
                  "attempted": 50,
                  "rate": 0.84,
                  "successes": 42,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 0.9166267863791231,
                    "low": 0.7148551903823401
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 8,
                  "route_failed": 0,
                  "route_unverifiable": 8,
                  "route_verified": 42,
                  "scheduled": 50,
                  "success": 42
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 42,
                      "ratio": 1.0,
                      "total": 42
                    },
                    "p50": 0.0015,
                    "p95": 0.00180975
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 42,
                      "ratio": 1.0,
                      "total": 42
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 42,
                      "ratio": 1.0,
                      "total": 42
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cold_end_to_end_first_body_byte_s": {
                    "coverage": {
                      "covered": 42,
                      "ratio": 1.0,
                      "total": 42
                    },
                    "p50": 0.8512310413643718,
                    "p95": 1.4127248435746877
                  },
                  "cold_end_to_end_response_headers_s": {
                    "coverage": {
                      "covered": 42,
                      "ratio": 1.0,
                      "total": 42
                    },
                    "p50": 0.847308375639841,
                    "p95": 1.4125103067141025
                  },
                  "cold_end_to_end_semantic_ttft_s": {
                    "coverage": {
                      "covered": 42,
                      "ratio": 1.0,
                      "total": 42
                    },
                    "p50": 1.5406583126168698,
                    "p95": 2.675047248369082
                  },
                  "cold_end_to_end_stream_total_s": {
                    "coverage": {
                      "covered": 42,
                      "ratio": 1.0,
                      "total": 42
                    },
                    "p50": 2.4771089998539537,
                    "p95": 3.6663599997526033
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 42,
                      "ratio": 1.0,
                      "total": 42
                    },
                    "p50": 43.0,
                    "p95": 45.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 42,
                      "ratio": 1.0,
                      "total": 42
                    },
                    "p50": 0.0015,
                    "p95": 0.00180975
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 42,
                      "ratio": 1.0,
                      "total": 42
                    },
                    "p50": 0.0015,
                    "p95": 0.00180975
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 42,
                      "ratio": 1.0,
                      "total": 42
                    },
                    "p50": 43.0,
                    "p95": 53.0
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 42,
                      "ratio": 1.0,
                      "total": 42
                    },
                    "p50": 2.2887690619099885,
                    "p95": 3.517570243636146
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 42,
                      "ratio": 1.0,
                      "total": 42
                    },
                    "p50": 0.7083673335146159,
                    "p95": 1.19752292681951
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 42,
                      "ratio": 1.0,
                      "total": 42
                    },
                    "p50": 0.7081590204034001,
                    "p95": 1.1965818774886423
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 42,
                      "ratio": 1.0,
                      "total": 42
                    },
                    "p50": 1.394085916923359,
                    "p95": 2.339439789485186
                  },
                  "setup_dns_s": {
                    "coverage": {
                      "covered": 42,
                      "ratio": 1.0,
                      "total": 42
                    },
                    "p50": 0.0021272499579936266,
                    "p95": 0.06012537851929639
                  },
                  "setup_tcp_s": {
                    "coverage": {
                      "covered": 42,
                      "ratio": 1.0,
                      "total": 42
                    },
                    "p50": 0.07232839614152908,
                    "p95": 0.21949114156886926
                  },
                  "setup_tls_s": {
                    "coverage": {
                      "covered": 42,
                      "ratio": 1.0,
                      "total": 42
                    },
                    "p50": 0.08157539553940296,
                    "p95": 0.22756288731470697
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 42,
                      "ratio": 1.0,
                      "total": 42
                    },
                    "p50": 51.50835572520252,
                    "p95": 63.73771078392612
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 42,
                      "ratio": 1.0,
                      "total": 42
                    },
                    "p50": 85.0,
                    "p95": 96.94999999999999
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 42,
                  "first_attempt_successes": 42,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 8
                }
              },
              "warm": {
                "availability": {
                  "attempted": 50,
                  "rate": 0.66,
                  "successes": 33,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 0.7756323036493894,
                    "low": 0.5215356854148286
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 17,
                  "route_failed": 0,
                  "route_unverifiable": 17,
                  "route_verified": 33,
                  "scheduled": 50,
                  "success": 33
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 33,
                      "ratio": 1.0,
                      "total": 33
                    },
                    "p50": 0.00297,
                    "p95": 0.023528
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 33,
                      "ratio": 1.0,
                      "total": 33
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 33,
                      "ratio": 1.0,
                      "total": 33
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 33,
                      "ratio": 1.0,
                      "total": 33
                    },
                    "p50": 43.0,
                    "p95": 45.4
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 33,
                      "ratio": 1.0,
                      "total": 33
                    },
                    "p50": 0.00151,
                    "p95": 0.001815
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 0.9090909090909091,
                      "total": 33
                    },
                    "p50": 0.002935,
                    "p95": 0.0032
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 33,
                      "ratio": 1.0,
                      "total": 33
                    },
                    "p50": 43.0,
                    "p95": 53.4
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 33,
                      "ratio": 1.0,
                      "total": 33
                    },
                    "p50": 1.9661719589494169,
                    "p95": 2.99518667506054
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 33,
                      "ratio": 1.0,
                      "total": 33
                    },
                    "p50": 0.5050150002352893,
                    "p95": 0.7437149078585207
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 33,
                      "ratio": 1.0,
                      "total": 33
                    },
                    "p50": 0.5048211254179478,
                    "p95": 0.7433067667298019
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 33,
                      "ratio": 1.0,
                      "total": 33
                    },
                    "p50": 1.0443799155764282,
                    "p95": 1.8167018503881978
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 33,
                      "ratio": 1.0,
                      "total": 33
                    },
                    "p50": 49.59424949421141,
                    "p95": 69.02643548786577
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 33,
                      "ratio": 1.0,
                      "total": 33
                    },
                    "p50": 86.0,
                    "p95": 96.0
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 45,
                    "2": 4,
                    "3": 1
                  },
                  "eventual_successes": 33,
                  "first_attempt_successes": 30,
                  "logical_cells": 50,
                  "retry_rescues": 3,
                  "unknown_cost_attempts": 23
                }
              }
            },
            "role": "gateway"
          },
          {
            "arm_id": "concentrate-openai",
            "baseline": false,
            "conditions": {
              "cold": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.001495,
                    "p95": 0.00180775
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cold_end_to_end_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.8144209794700146,
                    "p95": 1.1789401460438966
                  },
                  "cold_end_to_end_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.8142666663043201,
                    "p95": 1.1787753398995846
                  },
                  "cold_end_to_end_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.391483771149069,
                    "p95": 2.734516110108234
                  },
                  "cold_end_to_end_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.2923854789696634,
                    "p95": 3.8171396879246458
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 42.5,
                    "p95": 45.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.001495,
                    "p95": 0.00180775
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.001495,
                    "p95": 0.00180775
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 43.0,
                    "p95": 53.0
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.025058958446607,
                    "p95": 3.674110470898449
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.6257925210520625,
                    "p95": 0.9989466600818557
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.625650041969493,
                    "p95": 0.9987529229605568
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.1768357295077294,
                    "p95": 2.597582393768243
                  },
                  "setup_dns_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0019853960257023573,
                    "p95": 0.0035176751436665652
                  },
                  "setup_tcp_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.06759522901847959,
                    "p95": 0.20194963961839674
                  },
                  "setup_tls_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.082192687317729,
                    "p95": 0.21062039386015383
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 51.046006369474085,
                    "p95": 65.07346149502843
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 84.5,
                    "p95": 96.55
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              },
              "warm": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.003105,
                    "p95": 0.003417
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 42.0,
                    "p95": 45.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0015574999999999999,
                    "p95": 0.0017464999999999998
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.003105,
                    "p95": 0.003417
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 45.0,
                    "p95": 51.55
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.06865525059402,
                    "p95": 3.606683335849083
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.5573915834538639,
                    "p95": 0.9326671642949798
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.55727035459131,
                    "p95": 0.9324884019326412
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.1594057290349156,
                    "p95": 2.7132989773061107
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 49.68159912148096,
                    "p95": 64.21820004414757
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 87.5,
                    "p95": 93.55
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              }
            },
            "role": "gateway"
          },
          {
            "arm_id": "openrouter-openai",
            "baseline": false,
            "conditions": {
              "cold": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0015075000000000002,
                    "p95": 0.0018124999999999999
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 0,
                      "ratio": 0.0,
                      "total": 50
                    },
                    "p50": null,
                    "p95": null
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cold_end_to_end_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.6722005419433117,
                    "p95": 1.0246520979562774
                  },
                  "cold_end_to_end_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.5653555630706251,
                    "p95": 0.9949494874104855
                  },
                  "cold_end_to_end_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.260959229664877,
                    "p95": 2.0637487564003094
                  },
                  "cold_end_to_end_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.1196995829232037,
                    "p95": 3.022236775094643
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 42.5,
                    "p95": 45.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0015075000000000002,
                    "p95": 0.0018124999999999999
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0015075000000000002,
                    "p95": 0.0018124999999999999
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 43.0,
                    "p95": 53.099999999999994
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.8867723129224032,
                    "p95": 2.893937258375808
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.5029975830111653,
                    "p95": 0.8577915521571411
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.3752797085326165,
                    "p95": 0.7739186579361554
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.040494041517377,
                    "p95": 1.930915225134231
                  },
                  "setup_dns_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0019794164691120386,
                    "p95": 0.07830611274112012
                  },
                  "setup_tcp_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.06979091628454626,
                    "p95": 0.19299187501892445
                  },
                  "setup_tls_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.07953839632682502,
                    "p95": 0.19874338095542043
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 51.41394696385504,
                    "p95": 60.42425079518202
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 85.5,
                    "p95": 97.0
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              },
              "warm": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0030025,
                    "p95": 0.003457
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 0,
                      "ratio": 0.0,
                      "total": 50
                    },
                    "p50": null,
                    "p95": null
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 42.0,
                    "p95": 45.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.00151,
                    "p95": 0.0017705
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0030025,
                    "p95": 0.003457
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 43.5,
                    "p95": 52.0
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.8839431875385344,
                    "p95": 3.558356398181058
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.5137802921235561,
                    "p95": 0.7990115400636567
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.4100930420681834,
                    "p95": 0.7988973626866935
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.067598229041323,
                    "p95": 2.436416737479157
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 54.01470213757369,
                    "p95": 73.72451420180074
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 85.0,
                    "p95": 95.55
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              }
            },
            "role": "gateway"
          },
          {
            "arm_id": "ramp-openai",
            "baseline": false,
            "conditions": {
              "cold": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0014275,
                    "p95": 0.0017832500000000001
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cold_end_to_end_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.3313687711488456,
                    "p95": 6.457146087475118
                  },
                  "cold_end_to_end_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.331201208056882,
                    "p95": 6.4569749329239094
                  },
                  "cold_end_to_end_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.3324487290810794,
                    "p95": 6.458289426867847
                  },
                  "cold_end_to_end_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.802893521031365,
                    "p95": 7.075026627141049
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 42.0,
                    "p95": 45.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0014275,
                    "p95": 0.0017832500000000001
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0014275,
                    "p95": 0.0017832500000000001
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 41.0,
                    "p95": 52.55
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.7550697296392173,
                    "p95": 7.026515026972627
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.2677294376771897,
                    "p95": 6.408634487306696
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.2675618745852262,
                    "p95": 6.408463332755487
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.2688093956094235,
                    "p95": 6.4097778266994245
                  },
                  "setup_dns_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.002150937682017684,
                    "p95": 0.0025158205535262823
                  },
                  "setup_tcp_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.019747458631172776,
                    "p95": 0.036674370756372796
                  },
                  "setup_tls_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.02810522820800543,
                    "p95": 0.04262407114729285
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 37.51646969557622,
                    "p95": 67.13183119545906
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 83.5,
                    "p95": 94.1
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              },
              "warm": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0029375,
                    "p95": 0.00329375
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 42.0,
                    "p95": 45.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.001425,
                    "p95": 0.0017564999999999998
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0029375,
                    "p95": 0.00329375
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 40.5,
                    "p95": 51.099999999999994
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.9465111666359007,
                    "p95": 6.790144020807929
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.6926622295286506,
                    "p95": 6.2328684916021295
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.6924891041126102,
                    "p95": 6.232687968807294
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.6926622295286506,
                    "p95": 6.233270089607684
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 40.11674218758847,
                    "p95": 85.64563351396319
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 83.0,
                    "p95": 94.0
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              }
            },
            "role": "gateway",
            "score_baseline_arm": {
              "arm_id": "direct-openai",
              "baseline": true,
              "conditions": {
                "cold": {
                  "availability": {
                    "attempted": 50,
                    "rate": 1.0,
                    "successes": 50,
                    "wilson95": {
                      "confidence": 0.95,
                      "high": 1.0,
                      "low": 0.9286499658256813
                    }
                  },
                  "denominators": {
                    "attempted": 50,
                    "request_failed": 0,
                    "route_failed": 0,
                    "route_unverifiable": 0,
                    "route_verified": 50,
                    "scheduled": 50,
                    "success": 50
                  },
                  "metrics": {
                    "budget_debit_usd": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 0.0014550000000000001,
                      "p95": 0.0017737499999999997
                    },
                    "cache_write_input_tokens": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 0.0,
                      "p95": 0.0
                    },
                    "cached_input_tokens": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 0.0,
                      "p95": 0.0
                    },
                    "cold_end_to_end_first_body_byte_s": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 0.5404538328293711,
                      "p95": 1.3034997649490827
                    },
                    "cold_end_to_end_response_headers_s": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 0.5402706870809197,
                      "p95": 1.302919377549551
                    },
                    "cold_end_to_end_semantic_ttft_s": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 1.2048111038748175,
                      "p95": 2.0782052062917495
                    },
                    "cold_end_to_end_stream_total_s": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 2.091131979133934,
                      "p95": 3.3658654108177872
                    },
                    "input_tokens": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 42.0,
                      "p95": 45.0
                    },
                    "measured_cost_usd": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 0.0014550000000000001,
                      "p95": 0.0017737499999999997
                    },
                    "observed_cost_usd": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 0.0014550000000000001,
                      "p95": 0.0017737499999999997
                    },
                    "output_tokens": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 41.5,
                      "p95": 52.0
                    },
                    "request_stream_total_s": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 2.0489423545077443,
                      "p95": 3.314366120868362
                    },
                    "request_to_first_body_byte_s": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 0.49184256233274937,
                      "p95": 1.2552430083276698
                    },
                    "request_to_response_headers_s": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 0.491659416584298,
                      "p95": 1.254662620928138
                    },
                    "request_to_semantic_ttft_s": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 1.1309491673018783,
                      "p95": 2.015812387410551
                    },
                    "setup_dns_s": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 0.002478020265698433,
                      "p95": 0.040087527502328135
                    },
                    "setup_tcp_s": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 0.01909939548932016,
                      "p95": 0.025794085324741894
                    },
                    "setup_tls_s": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 0.02658360404893756,
                      "p95": 0.037893362599425
                    },
                    "throughput_tokens_per_s": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 52.8245826695139,
                      "p95": 69.16177203539158
                    },
                    "total_tokens": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 84.0,
                      "p95": 96.1
                    }
                  },
                  "retry_diagnostics": {
                    "attempt_count_distribution": {
                      "1": 50
                    },
                    "eventual_successes": 50,
                    "first_attempt_successes": 50,
                    "logical_cells": 50,
                    "retry_rescues": 0,
                    "unknown_cost_attempts": 0
                  }
                },
                "warm": {
                  "availability": {
                    "attempted": 50,
                    "rate": 1.0,
                    "successes": 50,
                    "wilson95": {
                      "confidence": 0.95,
                      "high": 1.0,
                      "low": 0.9286499658256813
                    }
                  },
                  "denominators": {
                    "attempted": 50,
                    "request_failed": 0,
                    "route_failed": 0,
                    "route_unverifiable": 0,
                    "route_verified": 50,
                    "scheduled": 50,
                    "success": 50
                  },
                  "metrics": {
                    "budget_debit_usd": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 0.0029575,
                      "p95": 0.0036047499999999994
                    },
                    "cache_write_input_tokens": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 0.0,
                      "p95": 0.0
                    },
                    "cached_input_tokens": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 0.0,
                      "p95": 0.0
                    },
                    "input_tokens": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 42.0,
                      "p95": 45.0
                    },
                    "measured_cost_usd": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 0.001445,
                      "p95": 0.0017455
                    },
                    "observed_cost_usd": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 0.0029575,
                      "p95": 0.0036047499999999994
                    },
                    "output_tokens": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 41.0,
                      "p95": 51.0
                    },
                    "request_stream_total_s": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 1.9241711250506341,
                      "p95": 2.7517991245491427
                    },
                    "request_to_first_body_byte_s": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 0.4260029373690486,
                      "p95": 0.8457103753229602
                    },
                    "request_to_response_headers_s": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 0.41929920809343457,
                      "p95": 0.8455722668673842
                    },
                    "request_to_semantic_ttft_s": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 1.0377994372975081,
                      "p95": 1.924336612760089
                    },
                    "throughput_tokens_per_s": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 54.14365228113792,
                      "p95": 73.77384349482251
                    },
                    "total_tokens": {
                      "coverage": {
                        "covered": 50,
                        "ratio": 1.0,
                        "total": 50
                      },
                      "p50": 84.0,
                      "p95": 94.1
                    }
                  },
                  "retry_diagnostics": {
                    "attempt_count_distribution": {
                      "1": 50
                    },
                    "eventual_successes": 50,
                    "first_attempt_successes": 50,
                    "logical_cells": 50,
                    "retry_rescues": 0,
                    "unknown_cost_attempts": 0
                  }
                }
              },
              "role": "direct"
            }
          },
          {
            "arm_id": "vercel-openai",
            "baseline": false,
            "conditions": {
              "cold": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0015025,
                    "p95": 0.0017004999999999998
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cold_end_to_end_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.27688706247136,
                    "p95": 2.7839966785162673
                  },
                  "cold_end_to_end_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.2670889794826508,
                    "p95": 2.7760851828847066
                  },
                  "cold_end_to_end_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.2783352294936776,
                    "p95": 2.785422731074503
                  },
                  "cold_end_to_end_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.195921478793025,
                    "p95": 3.557244422705843
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 42.5,
                    "p95": 45.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0015025,
                    "p95": 0.0017004999999999998
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0015025,
                    "p95": 0.0017004999999999998
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 43.0,
                    "p95": 49.0
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.970399541547522,
                    "p95": 3.294844472547993
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.0426163545344025,
                    "p95": 2.488366949744522
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.0188624795991927,
                    "p95": 2.4804554541129615
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.0439233335200697,
                    "p95": 2.4897930023027577
                  },
                  "setup_dns_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0030877708923071623,
                    "p95": 0.15704839809332044
                  },
                  "setup_tcp_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0707050203345716,
                    "p95": 0.20068323484156278
                  },
                  "setup_tls_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0892547289840877,
                    "p95": 0.20787209761328992
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 50.9114470385482,
                    "p95": 66.52310410603305
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 85.0,
                    "p95": 95.1
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              },
              "warm": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0029275,
                    "p95": 0.0033654999999999996
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 42.0,
                    "p95": 45.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0014675,
                    "p95": 0.0017104999999999998
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0029275,
                    "p95": 0.0033654999999999996
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 41.5,
                    "p95": 50.0
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.9116332295816392,
                    "p95": 3.7822423561010483
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.086522645317018,
                    "p95": 2.2954029770335174
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.0743594577070326,
                    "p95": 2.28886079811491
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.086522645317018,
                    "p95": 2.3858492082683362
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 51.31646077067815,
                    "p95": 69.04174459668917
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 84.0,
                    "p95": 93.0
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              }
            },
            "role": "gateway"
          }
        ],
        "baseline_arm_id": "direct-openai",
        "complete_blocks": {
          "cold": 50,
          "warm": 50
        },
        "completion_integrity": {
          "arms": {
            "cloudflare-openai": {
              "cold": {
                "measured": {
                  "length": 0,
                  "missing": 50,
                  "natural_stop": 0,
                  "other": 0,
                  "total": 50
                }
              },
              "warm": {
                "measured": {
                  "length": 0,
                  "missing": 50,
                  "natural_stop": 0,
                  "other": 0,
                  "total": 50
                },
                "warm_primer": {
                  "length": 0,
                  "missing": 50,
                  "natural_stop": 0,
                  "other": 0,
                  "total": 50
                }
              }
            },
            "concentrate-openai": {
              "cold": {
                "measured": {
                  "length": 0,
                  "missing": 50,
                  "natural_stop": 0,
                  "other": 0,
                  "total": 50
                }
              },
              "warm": {
                "measured": {
                  "length": 0,
                  "missing": 50,
                  "natural_stop": 0,
                  "other": 0,
                  "total": 50
                },
                "warm_primer": {
                  "length": 0,
                  "missing": 50,
                  "natural_stop": 0,
                  "other": 0,
                  "total": 50
                }
              }
            },
            "direct-openai": {
              "cold": {
                "measured": {
                  "length": 0,
                  "missing": 50,
                  "natural_stop": 0,
                  "other": 0,
                  "total": 50
                }
              },
              "warm": {
                "measured": {
                  "length": 0,
                  "missing": 50,
                  "natural_stop": 0,
                  "other": 0,
                  "total": 50
                },
                "warm_primer": {
                  "length": 0,
                  "missing": 50,
                  "natural_stop": 0,
                  "other": 0,
                  "total": 50
                }
              }
            },
            "openrouter-openai": {
              "cold": {
                "measured": {
                  "length": 0,
                  "missing": 50,
                  "natural_stop": 0,
                  "other": 0,
                  "total": 50
                }
              },
              "warm": {
                "measured": {
                  "length": 0,
                  "missing": 50,
                  "natural_stop": 0,
                  "other": 0,
                  "total": 50
                },
                "warm_primer": {
                  "length": 0,
                  "missing": 50,
                  "natural_stop": 0,
                  "other": 0,
                  "total": 50
                }
              }
            },
            "ramp-openai": {
              "cold": {
                "measured": {
                  "length": 0,
                  "missing": 50,
                  "natural_stop": 0,
                  "other": 0,
                  "total": 50
                }
              },
              "warm": {
                "measured": {
                  "length": 0,
                  "missing": 50,
                  "natural_stop": 0,
                  "other": 0,
                  "total": 50
                },
                "warm_primer": {
                  "length": 0,
                  "missing": 50,
                  "natural_stop": 0,
                  "other": 0,
                  "total": 50
                }
              }
            },
            "vercel-openai": {
              "cold": {
                "measured": {
                  "length": 0,
                  "missing": 50,
                  "natural_stop": 0,
                  "other": 0,
                  "total": 50
                }
              },
              "warm": {
                "measured": {
                  "length": 0,
                  "missing": 50,
                  "natural_stop": 0,
                  "other": 0,
                  "total": 50
                },
                "warm_primer": {
                  "length": 0,
                  "missing": 50,
                  "natural_stop": 0,
                  "other": 0,
                  "total": 50
                }
              }
            }
          }
        },
        "contrasts": [
          {
            "arm_id": "cloudflare-openai",
            "conditions": {
              "cold": {
                "budget_debit_usd": {
                  "coverage": {
                    "covered": 42,
                    "ratio": 0.84,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00014999999999999996,
                    "low": -3.000000000000008e-05,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 3.000000000000008e-05
                },
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 42,
                    "ratio": 0.84,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 42,
                    "ratio": 0.84,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cold_end_to_end_first_body_byte_s": {
                  "coverage": {
                    "covered": 42,
                    "ratio": 0.84,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.3334576250053942,
                    "low": 0.1629595209378749,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.2698553125374019
                },
                "cold_end_to_end_response_headers_s": {
                  "coverage": {
                    "covered": 42,
                    "ratio": 0.84,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.33067470882087946,
                    "low": 0.16241922904737294,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.26976818731054664
                },
                "cold_end_to_end_semantic_ttft_s": {
                  "coverage": {
                    "covered": 42,
                    "ratio": 0.84,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.36880199960432947,
                    "low": -0.009676541900262237,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.15970887546427548
                },
                "cold_end_to_end_stream_total_s": {
                  "coverage": {
                    "covered": 42,
                    "ratio": 0.84,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.7166268541477621,
                    "low": -0.0011123334988951683,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.2962817915249616
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 42,
                    "ratio": 0.84,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 42,
                    "ratio": 0.84,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00014999999999999996,
                    "low": -3.000000000000008e-05,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 3.000000000000008e-05
                },
                "observed_cost_usd": {
                  "coverage": {
                    "covered": 42,
                    "ratio": 0.84,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00014999999999999996,
                    "low": -3.000000000000008e-05,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 3.000000000000008e-05
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 42,
                    "ratio": 0.84,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 5.0,
                    "low": -1.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 1.0
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 42,
                    "ratio": 0.84,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.66422781185247,
                    "low": 0.014290313335368418,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.5099444794468582
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 42,
                    "ratio": 0.84,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.3916279789991677,
                    "low": 0.24055935442447662,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.3072123124729842
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 42,
                    "ratio": 0.84,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.3929703743197024,
                    "low": 0.24087531259283423,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.3074211459606886
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 42,
                    "ratio": 0.84,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.49424181249924004,
                    "low": -0.033486312720924616,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.17944435426034033
                },
                "setup_dns_s": {
                  "coverage": {
                    "covered": 42,
                    "ratio": 0.84,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -0.0008539790287613869,
                    "low": -0.07081520790234208,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.044626250164583325
                },
                "setup_tcp_s": {
                  "coverage": {
                    "covered": 42,
                    "ratio": 0.84,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.013715769397094846,
                    "low": -0.007490354124456644,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0009580841287970543
                },
                "setup_tls_s": {
                  "coverage": {
                    "covered": 42,
                    "ratio": 0.84,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.015212416416034102,
                    "low": -0.007594542112201452,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0008210206869989634
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 42,
                    "ratio": 0.84,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 6.971249500711416,
                    "low": -2.812199143472899,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 1.7134033899871248
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 42,
                    "ratio": 0.84,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 5.0,
                    "low": -1.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 1.0
                }
              },
              "warm": {
                "budget_debit_usd": {
                  "coverage": {
                    "covered": 33,
                    "ratio": 0.66,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 9.000000000000024e-05,
                    "low": -0.00011999999999999988,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -2.9999999999999645e-05
                },
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 33,
                    "ratio": 0.66,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 33,
                    "ratio": 0.66,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 33,
                    "ratio": 0.66,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 33,
                    "ratio": 0.66,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0001200000000000001,
                    "low": -2.9999999999999862e-05,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 5.999999999999994e-05
                },
                "observed_cost_usd": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 0.6,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 6.000000000000016e-05,
                    "low": -0.00014999999999999996,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -4.500000000000012e-05
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 33,
                    "ratio": 0.66,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 5.0,
                    "low": -1.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 2.0
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 33,
                    "ratio": 0.66,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.2619638745673001,
                    "low": -0.38548195734620094,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.06903495825827122
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 33,
                    "ratio": 0.66,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.22144416719675064,
                    "low": 0.1380630599334836,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.1566346655599773
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 33,
                    "ratio": 0.66,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.22136300103738904,
                    "low": 0.13954795757308602,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.1588059994392097
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 33,
                    "ratio": 0.66,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.19759329222142696,
                    "low": -0.01738149905577302,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.11638599867001176
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 33,
                    "ratio": 0.66,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 6.082529129437674,
                    "low": -0.3738032559985669,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 2.3469335531109152
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 33,
                    "ratio": 0.66,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 4.0,
                    "low": -1.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 2.0
                }
              }
            },
            "direct_arm": "direct-openai"
          },
          {
            "arm_id": "concentrate-openai",
            "conditions": {
              "cold": {
                "budget_debit_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00011999999999999999,
                    "low": -4.500000000000001e-05,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 6.000000000000005e-05
                },
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cold_end_to_end_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.2673949166201055,
                    "low": 0.10815135505981743,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.21639320789836347
                },
                "cold_end_to_end_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.26843033311888576,
                    "low": 0.10822593816556036,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.21635806234553456
                },
                "cold_end_to_end_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.17027066554874182,
                    "low": -0.17226551097701304,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.014809625223279
                },
                "cold_end_to_end_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.3912425832822919,
                    "low": -0.2387607293203473,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.15536743681877851
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0001200000000000001,
                    "low": -4.500000000000001e-05,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 6.000000000000005e-05
                },
                "observed_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00011999999999999999,
                    "low": -4.500000000000001e-05,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 6.000000000000005e-05
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 4.0,
                    "low": -1.5,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 2.0
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.35313800047151744,
                    "low": -0.15288743702694774,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.21927420794963837
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.2905651668552309,
                    "low": 0.18837116612121463,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.23351325001567602
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.2905818538274616,
                    "low": 0.18864376588608137,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.234395561972633
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.2556809578090906,
                    "low": -0.14925416675396264,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.07179454201832414
                },
                "setup_dns_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -0.000851999968290329,
                    "low": -0.07009487599134445,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.01075556268915534
                },
                "setup_tcp_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.009661873802542686,
                    "low": -0.008873812155798078,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.00027435459196567535
                },
                "setup_tls_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.013431083876639605,
                    "low": -0.003409250173717737,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.004593833349645138
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 4.255365440884692,
                    "low": -6.184841057369251,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.06383286755992046
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 4.0,
                    "low": -1.5,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 2.0
                }
              },
              "warm": {
                "budget_debit_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00021000000000000012,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0001200000000000001
                },
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0001200000000000001,
                    "low": -4.500000000000001e-05,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 1.4999999999999931e-05
                },
                "observed_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00021000000000000012,
                    "low": -1.4999999999999823e-05,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0001200000000000001
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 4.0,
                    "low": -1.5,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.5
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.39797408366575837,
                    "low": -0.12794852163642645,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.15085279080085456
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.2248150201048702,
                    "low": 0.1568608263041824,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.20514612505212426
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.223535334225744,
                    "low": 0.1572344198008068,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.2052416882943362
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.4334894160274416,
                    "low": -0.02941922890022397,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.16169824986718595
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 7.326570307074366,
                    "low": -3.049866731066185,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 2.0682682332649946
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 4.012499999999932,
                    "low": -1.5,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.5
                }
              }
            },
            "direct_arm": "direct-openai"
          },
          {
            "arm_id": "openrouter-openai",
            "conditions": {
              "cold": {
                "budget_debit_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 7.500000000000009e-05,
                    "low": -6.0375000000000024e-05,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 0,
                    "ratio": 0.0,
                    "total": 50
                  },
                  "interval": null,
                  "median_gateway_minus_direct": null
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cold_end_to_end_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.16925339596346003,
                    "low": 0.03226854186505079,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.09419237542897463
                },
                "cold_end_to_end_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.01860791677609086,
                    "low": -0.07030275127617641,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.010729084257036448
                },
                "cold_end_to_end_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -0.013774500228464603,
                    "low": -0.3240907702129334,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.1508644379209727
                },
                "cold_end_to_end_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.14753491594456136,
                    "low": -0.31320415415102615,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.1558502702973783
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 7.500000000000009e-05,
                    "low": -5.999999999999994e-05,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "observed_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 7.500000000000009e-05,
                    "low": -5.999999999999994e-05,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 2.5,
                    "low": -2.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.12245387467555702,
                    "low": -0.24872800009325147,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.11341066635213792
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.17817311484832304,
                    "low": 0.10989856231026351,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.1552668963558972
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.038997880846727624,
                    "low": -0.0035153753124177456,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.01868841634131968
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.06103158323094249,
                    "low": -0.25186145841144025,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.13897170848213136
                },
                "setup_dns_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -0.000669750152155757,
                    "low": -0.05969793791882694,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.0013141257222741842
                },
                "setup_tcp_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0013815828133374453,
                    "low": -0.010082739434437825,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.0035006660036742687
                },
                "setup_tls_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0036883604363538197,
                    "low": -0.004609666997566819,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.000461312010884285
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 3.1090943859998776,
                    "low": -4.187921782822556,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -2.040962205594351
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 2.0,
                    "low": -2.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                }
              },
              "warm": {
                "budget_debit_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00015000000000000018,
                    "low": -0.00011999999999999988,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 3.000000000000008e-05
                },
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 0,
                    "ratio": 0.0,
                    "total": 50
                  },
                  "interval": null,
                  "median_gateway_minus_direct": null
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00013500000000000003,
                    "low": -7.499999999999998e-05,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 1.4999999999999931e-05
                },
                "observed_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.000165,
                    "low": -0.00011999999999999988,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 3.000000000000008e-05
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 4.0,
                    "low": -2.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.5
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.25985475070774555,
                    "low": -0.2283213739283383,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.06747193727642298
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.1996543756686151,
                    "low": 0.1246528746560216,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.16382789448834956
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.08651072334614553,
                    "low": 0.014852332882583141,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.06149168708361685
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.16516314609907568,
                    "low": -0.12846604129299521,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.037990396143868566
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 9.85140367944334,
                    "low": -1.5919542416284926,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 2.550404344890218
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 4.5,
                    "low": -2.0124999999999993,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.5
                }
              }
            },
            "direct_arm": "direct-openai"
          },
          {
            "arm_id": "vercel-openai",
            "conditions": {
              "cold": {
                "budget_debit_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 5.999999999999994e-05,
                    "low": -3.000000000000008e-05,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 4.500000000000001e-05
                },
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cold_end_to_end_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.7718463956844062,
                    "low": 0.5262974789366126,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.6690576255787164
                },
                "cold_end_to_end_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.7792085830587894,
                    "low": 0.5155584164895117,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.6637110831215978
                },
                "cold_end_to_end_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.12521079229190946,
                    "low": -0.14129900094121695,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.05883102002553642
                },
                "cold_end_to_end_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.28122162469662726,
                    "low": -0.18717616703361273,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.020768916932865977
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 5.999999999999994e-05,
                    "low": -5.999999999999994e-05,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 4.500000000000001e-05
                },
                "observed_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 5.999999999999994e-05,
                    "low": -5.999999999999994e-05,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 4.500000000000001e-05
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 2.0,
                    "low": -2.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 1.5
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.20206425106152892,
                    "low": -0.11859235400334,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.021681833313778043
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.7525893542915583,
                    "low": 0.5840078750625253,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.63050552085042
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.7446436251047999,
                    "low": 0.5725892502814531,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.6130566457286477
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.08935279599972867,
                    "low": -0.21886902092956007,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.008458166383206844
                },
                "setup_dns_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0007288330234587193,
                    "low": -0.002065751742338751,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.00012131244875490665
                },
                "setup_tcp_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.013304583029821515,
                    "low": -0.00690458295866847,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.004669728921726346
                },
                "setup_tls_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.02469416568055749,
                    "low": 0.005163333844393492,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.013330374844372272
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 6.5118684329326575,
                    "low": -4.765313777544733,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 1.4840707104198358
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 2.0,
                    "low": -2.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 1.5
                }
              },
              "warm": {
                "budget_debit_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00018000000000000004,
                    "low": -0.00014999999999999996,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -4.4999999999999684e-05
                },
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 3.000000000000008e-05,
                    "low": -0.00011999999999999988,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "observed_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00018000000000000004,
                    "low": -0.00015000000000000018,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -4.4999999999999684e-05
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": -3.5,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.17039370816200972,
                    "low": -0.19351908401586115,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.01199681218713522
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.8627572921104729,
                    "low": 0.6462084981612861,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.7285688749980181
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.8548874580301344,
                    "low": 0.6408751653507352,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.7201293334364891
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.3077448750846088,
                    "low": -0.07955454220063984,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.08807491674087942
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 10.270545973130229,
                    "low": -0.05603584070031786,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 5.378436083458478
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": -4.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                }
              }
            },
            "direct_arm": "direct-openai"
          },
          {
            "arm_id": "ramp-openai",
            "conditions": {
              "cold": {
                "budget_debit_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 4.500000000000001e-05,
                    "low": -5.999999999999994e-05,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cold_end_to_end_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.33543383423239,
                    "low": 0.6163310005795211,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.8161516247782856
                },
                "cold_end_to_end_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.3354370421729982,
                    "low": 0.5826042089611292,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.8161298125050962
                },
                "cold_end_to_end_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.5774540631100535,
                    "low": 0.045760625042021275,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.25313329230993986
                },
                "cold_end_to_end_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.3831315622664988,
                    "low": 0.3630867078900337,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.8306385416071862
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 4.500000000000001e-05,
                    "low": -5.999999999999994e-05,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "observed_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 4.500000000000001e-05,
                    "low": -5.999999999999994e-05,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.0124999999999318,
                    "low": -2.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.369513832964003,
                    "low": 0.37796179112046957,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.8251857713330537
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.3357149166986346,
                    "low": 0.5807343546766788,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.8121064167935401
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.3357181246392429,
                    "low": 0.5838832301087677,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.8120846045203507
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.5419163759797812,
                    "low": 0.03620096854283478,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.2535228750202805
                },
                "setup_dns_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -0.00021666637621819973,
                    "low": -0.0005227699875831604,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.0003805209416896105
                },
                "setup_tcp_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0021877912804484367,
                    "low": -0.0011935620568692684,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0004559159278869629
                },
                "setup_tls_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.003180188126862049,
                    "low": 0.0006778540555387735,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0023489580489695072
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -8.171927131639809,
                    "low": -14.400789147577413,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -12.853599072740957
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.5,
                    "low": -2.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                }
              },
              "warm": {
                "budget_debit_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00014999999999999974,
                    "low": -0.00018000000000000004,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -5.999999999999994e-05
                },
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00011999999999999999,
                    "low": -6.000000000000016e-05,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 4.49999999999999e-05
                },
                "observed_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00014999999999999974,
                    "low": -0.00018000000000000004,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -5.999999999999994e-05
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 4.0,
                    "low": -2.5,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 1.5
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.5831876881420612,
                    "low": 0.5556532926857471,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.9083914794027805
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.5410301247611642,
                    "low": 0.8843627283349633,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 1.2946127504110336
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.5410598539747298,
                    "low": 0.8864973750896752,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 1.2946202922612429
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.8768466045148671,
                    "low": 0.27938676269259305,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.5458823749795556
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -0.32372087328075594,
                    "low": -17.576023598839562,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -14.47335964972813
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 4.0,
                    "low": -2.5,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 1.5
                }
              }
            },
            "direct_arm": "direct-openai"
          }
        ],
        "date": "2026-08-20",
        "default": true,
        "experiment_digest": "960f93e76135d177f74363ec9e4001378b92f5ef2f08b3561c377f3ee21a3e03",
        "experiment_id": "gpt-5.6-sol-five-way-responses-probe-credit-rerun",
        "experiment_path": "gateway-probe/2026-08-18-gpt56-sol-managed-100/experiment.json",
        "family": "gateway_probe",
        "id": "2026-08-18-gpt56-sol-managed-100",
        "kind": "release",
        "link": null,
        "model": "GPT-5.6 Sol",
        "model_match": "rolling_alias",
        "output_token_limit_equalities": {
          "arms": {
            "cloudflare-openai": {
              "cold": {
                "equal": 0,
                "measured": 42
              },
              "warm": {
                "equal": 0,
                "measured": 33
              }
            },
            "concentrate-openai": {
              "cold": {
                "equal": 0,
                "measured": 50
              },
              "warm": {
                "equal": 0,
                "measured": 50
              }
            },
            "direct-openai": {
              "cold": {
                "equal": 0,
                "measured": 50
              },
              "warm": {
                "equal": 0,
                "measured": 50
              }
            },
            "openrouter-openai": {
              "cold": {
                "equal": 0,
                "measured": 50
              },
              "warm": {
                "equal": 0,
                "measured": 50
              }
            },
            "ramp-openai": {
              "cold": {
                "equal": 0,
                "measured": 50
              },
              "warm": {
                "equal": 0,
                "measured": 50
              }
            },
            "vercel-openai": {
              "cold": {
                "equal": 0,
                "measured": 50
              },
              "warm": {
                "equal": 0,
                "measured": 50
              }
            }
          },
          "configured_limit": 256
        },
        "path": null,
        "price_digest": "ca00621f612b706ed9540e10d0e108b95bb60a5386cc0c0629a456cd8dbdb094",
        "result_count": 700,
        "result_schema_version": 4,
        "results_path": "gateway-probe/2026-08-18-gpt56-sol-managed-100/results.jsonl",
        "retry_summary": null,
        "run_note": "Five routes, each with 50 cold and 50 warm measured requests. Warm measurements include a separate primer. All routes use GPT-5.6 Sol with medium reasoning; temperature and top_p are omitted so provider defaults apply. Gateway routes lock to OpenAI with fallbacks and response caching disabled. Cloudflare returned HTTP 402 for 25 of 100 measured requests; availability is included in its composite score.",
        "schedule_digest": "2ca5f7b04e8860429aafe14365ee02f230c9cf96dc81644f1860c9dd47ec1463",
        "scheduled_blocks_per_condition": 50,
        "submitter": "OpenBench",
        "supplemental_runs": [
          {
            "arm_ids": [
              "ramp-openai"
            ],
            "date": "2026-08-20",
            "experiment_digest": "86d9f9d746246fdde56b3ea2149913c057955a3cb3c367440f71732df545efe5",
            "experiment_path": "gateway-probe/2026-08-20-gpt56-ramp-supplemental-100/experiment.json",
            "id": "2026-08-20-gpt56-ramp-supplemental-100",
            "results_path": "gateway-probe/2026-08-20-gpt56-ramp-supplemental-100/results.jsonl",
            "run_note": "Ramp Router adds 50 cold and 50 warm measured requests collected on 2026-08-20. Its direct-relative deltas use the fresh direct OpenAI control from the same supplemental run, not the 2026-08-18 control. Ramp returns the served GPT-5.6 Sol model ID but not an upstream-provider field; OpenBench derives OpenAI ownership from the admitted model identity.",
            "verified_with_commit": "b7606d1f450f655e41a5ed28b7f890342941093b"
          }
        ],
        "title": "GPT-5.6 Sol Gateway Bench, 100 requests per route",
        "track": "request_probe",
        "verified_with_commit": "73c569ebf06a67e4935e5e850a827ebb7d3c78a9"
      },
      {
        "arms": [
          {
            "arm_id": "direct-deepseek",
            "baseline": true,
            "conditions": {
              "cold": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.5410000000000002e-05,
                    "p95": 2.954e-05
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 0,
                      "ratio": 0.0,
                      "total": 50
                    },
                    "p50": null,
                    "p95": null
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cold_end_to_end_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.3789753125165589,
                    "p95": 0.4600321706355317
                  },
                  "cold_end_to_end_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.3711659164982848,
                    "p95": 0.45983618923928593
                  },
                  "cold_end_to_end_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.8758773125009611,
                    "p95": 1.1715591687418054
                  },
                  "cold_end_to_end_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.299438937479863,
                    "p95": 1.6197765813936709
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 116.0,
                    "p95": 119.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.5410000000000002e-05,
                    "p95": 2.954e-05
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.5410000000000002e-05,
                    "p95": 2.954e-05
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 33.0,
                    "p95": 47.0
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.271555500512477,
                    "p95": 1.6009910733875583
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.35607547950348817,
                    "p95": 0.4244687940343283
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.3453630209842231,
                    "p95": 0.4243379753024783
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.8533482499769889,
                    "p95": 1.1439464607596164
                  },
                  "setup_dns_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.001880333002191037,
                    "p95": 0.031213781551923586
                  },
                  "setup_tcp_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.006207667000126094,
                    "p95": 0.007471591842477209
                  },
                  "setup_tls_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.009183582995319739,
                    "p95": 0.012673958527739157
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 79.61617573268744,
                    "p95": 97.31290549891092
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 149.5,
                    "p95": 163.1
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              },
              "warm": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 5.068e-05,
                    "p95": 6.831999999999998e-05
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 0,
                      "ratio": 0.0,
                      "total": 50
                    },
                    "p50": null,
                    "p95": null
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 116.0,
                    "p95": 118.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.5129999999999998e-05,
                    "p95": 3.1717e-05
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 5.068e-05,
                    "p95": 6.831999999999998e-05
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 32.0,
                    "p95": 55.099999999999994
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.273796207999112,
                    "p95": 1.62951974972093
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.3461807294806931,
                    "p95": 0.42392302966036366
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.3378102295100689,
                    "p95": 0.4168035459064413
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.8635460624936968,
                    "p95": 1.1534000350482527
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 81.77405271852298,
                    "p95": 102.46177973206898
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 147.5,
                    "p95": 171.55
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              }
            },
            "role": "direct"
          },
          {
            "arm_id": "cloudflare-deepseek",
            "baseline": false,
            "conditions": {
              "cold": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.8129999999999998e-05,
                    "p95": 2.1664999999999997e-05
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 0,
                      "ratio": 0.0,
                      "total": 50
                    },
                    "p50": null,
                    "p95": null
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cold_end_to_end_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.0781583749921992,
                    "p95": 4.267010693770132
                  },
                  "cold_end_to_end_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.077627291000681,
                    "p95": 4.266125112515871
                  },
                  "cold_end_to_end_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.0782848540111445,
                    "p95": 4.267193206262888
                  },
                  "cold_end_to_end_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.65860102098668,
                    "p95": 5.18601890627178
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 37.0,
                    "p95": 40.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.8129999999999998e-05,
                    "p95": 2.1664999999999997e-05
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.8129999999999998e-05,
                    "p95": 2.1664999999999997e-05
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 46.5,
                    "p95": 59.099999999999994
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.618868437508354,
                    "p95": 5.1564573496958435
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.0217982294852845,
                    "p95": 4.238055099995105
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.0216349374968559,
                    "p95": 4.237169518740845
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.022066312521929,
                    "p95": 4.238237612487861
                  },
                  "setup_dns_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.001496854005381465,
                    "p95": 0.020927083457354452
                  },
                  "setup_tcp_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.011596958473091945,
                    "p95": 0.012706145903212018
                  },
                  "setup_tls_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0156009795027785,
                    "p95": 0.019236070822807958
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 100.77721262668176,
                    "p95": 352.3384433299914
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 83.5,
                    "p95": 96.55
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              },
              "warm": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 3.6610000000000004e-05,
                    "p95": 4.8216e-05
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 0,
                      "ratio": 0.0,
                      "total": 50
                    },
                    "p50": null,
                    "p95": null
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 37.0,
                    "p95": 39.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.778e-05,
                    "p95": 2.8118999999999996e-05
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 3.6610000000000004e-05,
                    "p95": 4.8216e-05
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 45.0,
                    "p95": 81.64999999999999
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.664803521009162,
                    "p95": 5.137605431399424
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.001635416498175,
                    "p95": 2.792440454513416
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.0015353744965978,
                    "p95": 2.7923203463695234
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.0018291249871254,
                    "p95": 2.7926031374896403
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 85.17392466626212,
                    "p95": 306.5945316936511
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 82.0,
                    "p95": 119.19999999999999
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              }
            },
            "role": "gateway"
          },
          {
            "arm_id": "concentrate-deepseek",
            "baseline": false,
            "conditions": {
              "cold": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.52e-05,
                    "p95": 2.9008e-05
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cold_end_to_end_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.5883609369629994,
                    "p95": 0.7663224733929384
                  },
                  "cold_end_to_end_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.5882514999830164,
                    "p95": 0.7661965777253499
                  },
                  "cold_end_to_end_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.1326697289769072,
                    "p95": 1.5729301348968874
                  },
                  "cold_end_to_end_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.5186781669908669,
                    "p95": 2.0457793210516684
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 116.0,
                    "p95": 119.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.52e-05,
                    "p95": 2.9008e-05
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.52e-05,
                    "p95": 2.9008e-05
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 32.0,
                    "p95": 45.64999999999999
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.4905653129681014,
                    "p95": 2.0141451856121417
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.5579639579809736,
                    "p95": 0.7350186190684326
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.5578036664810497,
                    "p95": 0.7348655919922749
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.1045009164954536,
                    "p95": 1.5390500517300096
                  },
                  "setup_dns_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0013102914672344923,
                    "p95": 0.001806283285259269
                  },
                  "setup_tcp_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.011520833504619077,
                    "p95": 0.013605712502612732
                  },
                  "setup_tls_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.01706479149288498,
                    "p95": 0.021995668447925706
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 80.77303019718164,
                    "p95": 103.05138747665605
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 148.0,
                    "p95": 162.55
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              },
              "warm": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 5.0610000000000005e-05,
                    "p95": 5.8037e-05
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 116.0,
                    "p95": 118.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.52e-05,
                    "p95": 2.795099999999999e-05
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 5.0610000000000005e-05,
                    "p95": 5.8037e-05
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 32.0,
                    "p95": 42.04999999999997
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.4262051040132064,
                    "p95": 1.680484712330508
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.47827489601331763,
                    "p95": 0.6030239935847931
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.4781843335076701,
                    "p95": 0.6029013104591285
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.0355864165176172,
                    "p95": 1.2623948269028915
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 80.62438477717133,
                    "p95": 105.19951166941377
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 148.0,
                    "p95": 158.04999999999995
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              }
            },
            "role": "gateway"
          },
          {
            "arm_id": "openrouter-deepseek",
            "baseline": false,
            "conditions": {
              "cold": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.52e-05,
                    "p95": 3.1717e-05
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cold_end_to_end_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.5544545420270879,
                    "p95": 0.6060446418588981
                  },
                  "cold_end_to_end_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.47943795900209807,
                    "p95": 0.5844119500281522
                  },
                  "cold_end_to_end_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.0101946250069886,
                    "p95": 1.2393757294950773
                  },
                  "cold_end_to_end_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.444859874987742,
                    "p95": 1.7871203604445327
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 116.0,
                    "p95": 119.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.52e-05,
                    "p95": 3.1717e-05
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.52e-05,
                    "p95": 3.1717e-05
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 32.0,
                    "p95": 55.0
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.4179722294793464,
                    "p95": 1.7598203062661923
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.5221784794994164,
                    "p95": 0.5786327088630059
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.4462737290014047,
                    "p95": 0.5564564981352305
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.9781747914967127,
                    "p95": 1.2122898001485738
                  },
                  "setup_dns_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0013421459880191833,
                    "p95": 0.02051179186091758
                  },
                  "setup_tcp_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.01141764601925388,
                    "p95": 0.014245139242848377
                  },
                  "setup_tls_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.014898937486577779,
                    "p95": 0.019190104480367155
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 76.93781049422708,
                    "p95": 103.12228941096414
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 147.5,
                    "p95": 171.55
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              },
              "warm": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 5.075e-05,
                    "p95": 6.0675999999999986e-05
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 116.0,
                    "p95": 118.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.5410000000000002e-05,
                    "p95": 3.3082e-05
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 5.075e-05,
                    "p95": 6.0675999999999986e-05
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 33.0,
                    "p95": 59.64999999999999
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.389982978522312,
                    "p95": 1.6526284892461263
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.46202922900556587,
                    "p95": 0.4798857434594538
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.3785679585125763,
                    "p95": 0.4772026730701327
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.9173715835204348,
                    "p95": 1.1724540831462946
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 78.08311091416954,
                    "p95": 105.80112019558584
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 149.0,
                    "p95": 176.64999999999998
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              }
            },
            "role": "gateway"
          },
          {
            "arm_id": "vercel-deepseek",
            "baseline": false,
            "conditions": {
              "cold": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.527e-05,
                    "p95": 3.0737e-05
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 0,
                      "ratio": 0.0,
                      "total": 50
                    },
                    "p50": null,
                    "p95": null
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cold_end_to_end_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.5454867085209116,
                    "p95": 0.663545751926722
                  },
                  "cold_end_to_end_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.5410922294831835,
                    "p95": 0.6582183602062287
                  },
                  "cold_end_to_end_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.5946970210352447,
                    "p95": 6.869229941695808
                  },
                  "cold_end_to_end_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.004770521016326,
                    "p95": 7.16648341043037
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 116.0,
                    "p95": 119.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.527e-05,
                    "p95": 3.0737e-05
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.527e-05,
                    "p95": 3.0737e-05
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 32.0,
                    "p95": 52.0
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.9599485204962548,
                    "p95": 7.1354530001350165
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.5128176870057359,
                    "p95": 0.6314988353435182
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.5019514584855642,
                    "p95": 0.6259476812265348
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.562633416528115,
                    "p95": 6.8381995314004556
                  },
                  "setup_dns_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.001749541494064033,
                    "p95": 0.043409481382695916
                  },
                  "setup_tcp_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.005605333513813093,
                    "p95": 0.007968504179734735
                  },
                  "setup_tls_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.022980000008828938,
                    "p95": 0.026921395491808654
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 81.90535370578273,
                    "p95": 106.90891134034547
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 149.0,
                    "p95": 167.55
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              },
              "warm": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 5.04e-05,
                    "p95": 6.2517e-05
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 0,
                      "ratio": 0.0,
                      "total": 50
                    },
                    "p50": null,
                    "p95": null
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 116.0,
                    "p95": 118.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.492e-05,
                    "p95": 3.0141999999999997e-05
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 5.04e-05,
                    "p95": 6.2517e-05
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 31.0,
                    "p95": 50.749999999999986
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.785707145521883,
                    "p95": 9.123225145513423
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.437427041499177,
                    "p95": 0.5305207918630912
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.4282626874919515,
                    "p95": 0.5265792309684911
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 1.328305208531674,
                    "p95": 8.705567266675644
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 81.59526483448077,
                    "p95": 187.62022739611803
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 147.5,
                    "p95": 164.55
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              }
            },
            "role": "gateway"
          }
        ],
        "baseline_arm_id": "direct-deepseek",
        "complete_blocks": {
          "cold": 50,
          "warm": 50
        },
        "completion_integrity": {
          "arms": {
            "cloudflare-deepseek": {
              "cold": {
                "measured": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                }
              },
              "warm": {
                "measured": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                },
                "warm_primer": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                }
              }
            },
            "concentrate-deepseek": {
              "cold": {
                "measured": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                }
              },
              "warm": {
                "measured": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                },
                "warm_primer": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                }
              }
            },
            "direct-deepseek": {
              "cold": {
                "measured": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                }
              },
              "warm": {
                "measured": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                },
                "warm_primer": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                }
              }
            },
            "openrouter-deepseek": {
              "cold": {
                "measured": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                }
              },
              "warm": {
                "measured": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                },
                "warm_primer": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                }
              }
            },
            "vercel-deepseek": {
              "cold": {
                "measured": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                }
              },
              "warm": {
                "measured": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                },
                "warm_primer": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                }
              }
            }
          }
        },
        "contrasts": [
          {
            "arm_id": "cloudflare-deepseek",
            "conditions": {
              "cold": {
                "budget_debit_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -6.999999999999999e-06,
                    "low": -9.523499999999998e-06,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -8.54e-06
                },
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 0,
                    "ratio": 0.0,
                    "total": 50
                  },
                  "interval": null,
                  "median_gateway_minus_direct": null
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cold_end_to_end_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.8188752294809092,
                    "low": 0.5691181249567308,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.6686293329985347
                },
                "cold_end_to_end_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.8369722079951316,
                    "low": 0.5762028750032187,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.6754218959831633
                },
                "cold_end_to_end_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.35498112498316914,
                    "low": 0.04171491652959958,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.15821179150952958
                },
                "cold_end_to_end_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.4784390415006783,
                    "low": 0.07374583400087431,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.23148547901655547
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -79.0,
                    "low": -79.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -79.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -6.999999999999999e-06,
                    "low": -9.659999999999999e-06,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -8.54e-06
                },
                "observed_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -6.999999999999999e-06,
                    "low": -9.519999999999999e-06,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -8.54e-06
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 14.5,
                    "low": 5.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 9.0
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.4633332920493558,
                    "low": 0.06212617814453552,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.2209621244983282
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.8359836244781036,
                    "low": 0.5695528534706682,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.6734084794879891
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.8112262284848839,
                    "low": 0.5818853165401379,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.6757151460333262
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.35694245802005753,
                    "low": 0.05325235353666358,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.1447508950077463
                },
                "setup_dns_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -0.0002353124727960676,
                    "low": -0.0008488330349791795,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.0004051670548506081
                },
                "setup_tcp_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0055223749950528145,
                    "low": 0.005077040987089276,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0053511875157710165
                },
                "setup_tls_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.006906583468662575,
                    "low": 0.005812166986288503,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.006183187011629343
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 36.87252097172447,
                    "low": 9.94829944607271,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 19.315539403559477
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -64.5,
                    "low": -74.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -70.0
                }
              },
              "warm": {
                "budget_debit_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -1.1199999999999998e-05,
                    "low": -1.6519999999999998e-05,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -1.414e-05
                },
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 0,
                    "ratio": 0.0,
                    "total": 50
                  },
                  "interval": null,
                  "median_gateway_minus_direct": null
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -79.0,
                    "low": -79.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -79.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -6.02e-06,
                    "low": -9.1e-06,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -7.2800000000000014e-06
                },
                "observed_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -1.1479999999999999e-05,
                    "low": -1.6519999999999998e-05,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -1.414e-05
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 18.0,
                    "low": 7.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 13.5
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.5538244159542955,
                    "low": 0.20263154100393876,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.3741921664914116
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.9054478330072016,
                    "low": 0.4569131670286879,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.6603136454650667
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.8645987084892113,
                    "low": 0.4592857910320163,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.6681887700106017
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.3013449575228151,
                    "low": -0.04712595799355768,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.16785754149896093
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 22.576458283710128,
                    "low": -10.54574891012281,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.7342646882629467
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -61.0,
                    "low": -72.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -65.5
                }
              }
            },
            "direct_arm": "direct-deepseek"
          },
          {
            "arm_id": "concentrate-deepseek",
            "conditions": {
              "cold": {
                "budget_debit_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 5.600000000000014e-07,
                    "low": -8.400000000000021e-07,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 0,
                    "ratio": 0.0,
                    "total": 50
                  },
                  "interval": null,
                  "median_gateway_minus_direct": null
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cold_end_to_end_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.23617195853148587,
                    "low": 0.18397867384410346,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.1980972710007336
                },
                "cold_end_to_end_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.22984106300282292,
                    "low": 0.18869583396008238,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.20513854100136086
                },
                "cold_end_to_end_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.3616938960039988,
                    "low": 0.16098991702892818,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.24072252050973475
                },
                "cold_end_to_end_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.31805612496100366,
                    "low": 0.1224918125080876,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.24020954201114364
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 5.600000000000014e-07,
                    "low": -8.400000000000021e-07,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "observed_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 5.600000000000014e-07,
                    "low": -8.400000000000021e-07,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 2.0,
                    "low": -3.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.3124894296182901,
                    "low": 0.1218967919703573,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.24859820798155852
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.22361791599541903,
                    "low": 0.17787512595532462,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.1977867710229475
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.22362733201589435,
                    "low": 0.18461924901930615,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.20304922954528593
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.34999383403919637,
                    "low": 0.15888812454068102,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.2478514580288902
                },
                "setup_dns_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -0.0004285000031813979,
                    "low": -0.0008261460170615464,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.0005340625066310167
                },
                "setup_tcp_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00556254101684317,
                    "low": 0.004909020499326289,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.00517635396681726
                },
                "setup_tls_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.008131124486681074,
                    "low": 0.006966292014112696,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.007650687533896416
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 4.881900544286261,
                    "low": -7.020808068709627,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -2.1000313773890866
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 2.0,
                    "low": -3.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                }
              },
              "warm": {
                "budget_debit_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 5.600000000000014e-07,
                    "low": -1.4000000000000035e-06,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -4.1999999999999937e-07
                },
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 0,
                    "ratio": 0.0,
                    "total": 50
                  },
                  "interval": null,
                  "median_gateway_minus_direct": null
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 2.800000000000007e-07,
                    "low": -7.000000000000001e-07,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -2.800000000000007e-07
                },
                "observed_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 5.600000000000014e-07,
                    "low": -1.4000000000000001e-06,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -4.1999999999999937e-07
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": -2.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -1.0
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.2265718340058811,
                    "low": 0.0723076249123551,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.15893072896869853
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.14839550002943724,
                    "low": 0.11586683202767745,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.13617431247257628
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.15599487401777878,
                    "low": 0.12277022947091609,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.142840229003923
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.21835374998045154,
                    "low": 0.07560746931412723,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.17202089601778425
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 5.439440439759505,
                    "low": -5.181572129391682,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -1.9188531476346498
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": -2.5,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -1.0
                }
              }
            },
            "direct_arm": "direct-deepseek"
          },
          {
            "arm_id": "openrouter-deepseek",
            "conditions": {
              "cold": {
                "budget_debit_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 8.400000000000004e-07,
                    "low": -5.600000000000014e-07,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -1.3999999999999866e-07
                },
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 0,
                    "ratio": 0.0,
                    "total": 50
                  },
                  "interval": null,
                  "median_gateway_minus_direct": null
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cold_end_to_end_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.18033233367750653,
                    "low": 0.16320008301408961,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.17137620798894204
                },
                "cold_end_to_end_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.11891900000046007,
                    "low": 0.09316516597755253,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.10919487496721558
                },
                "cold_end_to_end_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.21387193800183013,
                    "low": 0.07760669977869838,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.13628100004279986
                },
                "cold_end_to_end_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.24101112404605374,
                    "low": 0.020651064083358504,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.1607440625084564
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 8.400000000000004e-07,
                    "low": -5.600000000000014e-07,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -1.3999999999999866e-07
                },
                "observed_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 8.399999999999988e-07,
                    "low": -5.600000000000014e-07,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -1.3999999999999866e-07
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 3.0,
                    "low": -2.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.5
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.24199527100427076,
                    "low": 0.02388512500328943,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.16861506254645064
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.17977729300037026,
                    "low": 0.15110187500249594,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.16433281198260374
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.11223814340701208,
                    "low": 0.08192404097644612,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.09755575048620813
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.21637926998664625,
                    "low": 0.073255583003629,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.1287000625161454
                },
                "setup_dns_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -0.00027554098051041365,
                    "low": -0.0008508329920005053,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.0005356664769351482
                },
                "setup_tcp_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.005303499987348914,
                    "low": 0.004920728999422863,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0051207710057497025
                },
                "setup_tls_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0062780209991615266,
                    "low": 0.005269832967314869,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.005827999528264627
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.1494297503893605,
                    "low": -7.671305962180554,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -3.200985161572973
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 3.0,
                    "low": -2.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.5
                }
              },
              "warm": {
                "budget_debit_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.5399999999999988e-06,
                    "low": -8.399999999999987e-07,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 0,
                    "ratio": 0.0,
                    "total": 50
                  },
                  "interval": null,
                  "median_gateway_minus_direct": null
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 8.399999999999987e-07,
                    "low": -1.3999999999999866e-07,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "observed_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.4000000000000001e-06,
                    "low": -8.399999999999987e-07,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 3.0,
                    "low": -0.5,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.19340495800133795,
                    "low": -0.04527797899208963,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.12833747998229228
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.12186212599044666,
                    "low": 0.10845170903485268,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.11746116649010219
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.04969093750696629,
                    "low": 0.03209775054710917,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.03848768802708946
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.11330029144301079,
                    "low": -0.03769000101601705,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.058347541984403506
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 3.0936031563230344,
                    "low": -11.568924015680409,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -4.876710421637121
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 3.0,
                    "low": -0.01249999999999929,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                }
              }
            },
            "direct_arm": "direct-deepseek"
          },
          {
            "arm_id": "vercel-deepseek",
            "conditions": {
              "cold": {
                "budget_debit_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 5.600000000000014e-07,
                    "low": -8.400000000000004e-07,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 0,
                    "ratio": 0.0,
                    "total": 50
                  },
                  "interval": null,
                  "median_gateway_minus_direct": null
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cold_end_to_end_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.1741267914767377,
                    "low": 0.1449657499906607,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.15299602152663283
                },
                "cold_end_to_end_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.18204187497030944,
                    "low": 0.1420199170242995,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.1528207714727614
                },
                "cold_end_to_end_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.3076719589880668,
                    "low": 0.2925302666888456,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.726469667017227
                },
                "cold_end_to_end_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.4403943340294063,
                    "low": 0.3859690419922117,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.7033316045126412
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 7.000000000000001e-07,
                    "low": -8.400000000000004e-07,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "observed_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 7.000000000000001e-07,
                    "low": -8.400000000000004e-07,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 2.5,
                    "low": -3.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.4248190829530358,
                    "low": 0.40908216229727257,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.6943704790028278
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.15198508294997737,
                    "low": 0.12508339603664353,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.13491552052437328
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.16043358301976696,
                    "low": 0.12060360453324392,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.13231437496142462
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.3176874999771826,
                    "low": 0.28638800006592646,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.7119829994917382
                },
                "setup_dns_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 8.066598093137145e-05,
                    "low": -0.0003587090177461505,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -9.166699601337314e-05
                },
                "setup_tcp_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -0.00017799995839595795,
                    "low": -0.0010445010557305068,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.0006634995224885643
                },
                "setup_tls_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.014395749982213601,
                    "low": 0.013016165990848094,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.014011042047059163
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 8.807437954827392,
                    "low": -4.341302536924999,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 2.7762479475313384
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 2.5,
                    "low": -3.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                }
              },
              "warm": {
                "budget_debit_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 8.399999999999987e-07,
                    "low": -1.6799999999999975e-06,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -5.600000000000014e-07
                },
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 0,
                    "ratio": 0.0,
                    "total": 50
                  },
                  "interval": null,
                  "median_gateway_minus_direct": null
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": -8.399999999999987e-07,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -2.800000000000007e-07
                },
                "observed_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 8.399999999999987e-07,
                    "low": -1.6799999999999975e-06,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -5.600000000000014e-07
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": -3.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -1.0
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.1712328540161252,
                    "low": 0.20652424998115748,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.5091008745075669
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.11041469269403015,
                    "low": 0.0698564994672779,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0872749995032791
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.1086736680008471,
                    "low": 0.06919520901283249,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.08752627047942951
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.228858999995282,
                    "low": 0.22559345717236282,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.44528404195443727
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 7.485737502561065,
                    "low": -8.113782169501903,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.8772755223536208
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": -3.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -1.0
                }
              }
            },
            "direct_arm": "direct-deepseek"
          }
        ],
        "date": "2026-08-03",
        "default": false,
        "experiment_digest": "34873bc948a088c2a5e23b3a356ed3425f71418e80f6f14afd400f6b73dcd8c4",
        "experiment_id": "deepseek-v4-flash-five-way-chat-probe",
        "experiment_path": "gateway-probe/2026-08-03-deepseek-v4-flash-100/experiment.json",
        "family": "gateway_probe",
        "id": "2026-08-03-deepseek-v4-flash-100",
        "kind": "release",
        "link": null,
        "model": "DeepSeek V4 Flash",
        "model_match": "rolling_alias",
        "output_token_limit_equalities": {
          "arms": {
            "cloudflare-deepseek": {
              "cold": {
                "equal": 0,
                "measured": 50
              },
              "warm": {
                "equal": 0,
                "measured": 50
              }
            },
            "concentrate-deepseek": {
              "cold": {
                "equal": 0,
                "measured": 50
              },
              "warm": {
                "equal": 0,
                "measured": 50
              }
            },
            "direct-deepseek": {
              "cold": {
                "equal": 0,
                "measured": 50
              },
              "warm": {
                "equal": 0,
                "measured": 50
              }
            },
            "openrouter-deepseek": {
              "cold": {
                "equal": 0,
                "measured": 50
              },
              "warm": {
                "equal": 0,
                "measured": 50
              }
            },
            "vercel-deepseek": {
              "cold": {
                "equal": 0,
                "measured": 50
              },
              "warm": {
                "equal": 0,
                "measured": 50
              }
            }
          },
          "configured_limit": 4096
        },
        "path": null,
        "price_digest": "4fbc8bcc47b18dfe93015fe17a5c2d87dc1455d8a192ea8d68073d59e505e93a",
        "result_count": 500,
        "result_schema_version": 4,
        "results_path": "gateway-probe/2026-08-03-deepseek-v4-flash-100/results.jsonl",
        "retry_summary": {
          "groups": [
            {
              "arm_id": "cloudflare-deepseek",
              "attempt_distribution": {
                "1": 50
              },
              "attempts": 50,
              "condition": "cold",
              "eventual_successes": 50,
              "exhausted": 0,
              "final_attempt_start_offset_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "first_attempt_successes": 50,
              "logical_cells": 50,
              "output_limits": {
                "4096": {
                  "equal": 0,
                  "measured": 50
                }
              },
              "recovery_completion_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "rescued": 0,
              "retried": 0,
              "retry_after_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "wait_actual_s": {
                "count": 0,
                "max": null,
                "median": null
              }
            },
            {
              "arm_id": "cloudflare-deepseek",
              "attempt_distribution": {
                "1": 50
              },
              "attempts": 50,
              "condition": "warm",
              "eventual_successes": 50,
              "exhausted": 0,
              "final_attempt_start_offset_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "first_attempt_successes": 50,
              "logical_cells": 50,
              "output_limits": {
                "4096": {
                  "equal": 0,
                  "measured": 50
                }
              },
              "recovery_completion_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "rescued": 0,
              "retried": 0,
              "retry_after_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "wait_actual_s": {
                "count": 0,
                "max": null,
                "median": null
              }
            },
            {
              "arm_id": "concentrate-deepseek",
              "attempt_distribution": {
                "1": 50
              },
              "attempts": 50,
              "condition": "cold",
              "eventual_successes": 50,
              "exhausted": 0,
              "final_attempt_start_offset_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "first_attempt_successes": 50,
              "logical_cells": 50,
              "output_limits": {
                "4096": {
                  "equal": 0,
                  "measured": 50
                }
              },
              "recovery_completion_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "rescued": 0,
              "retried": 0,
              "retry_after_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "wait_actual_s": {
                "count": 0,
                "max": null,
                "median": null
              }
            },
            {
              "arm_id": "concentrate-deepseek",
              "attempt_distribution": {
                "1": 50
              },
              "attempts": 50,
              "condition": "warm",
              "eventual_successes": 50,
              "exhausted": 0,
              "final_attempt_start_offset_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "first_attempt_successes": 50,
              "logical_cells": 50,
              "output_limits": {
                "4096": {
                  "equal": 0,
                  "measured": 50
                }
              },
              "recovery_completion_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "rescued": 0,
              "retried": 0,
              "retry_after_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "wait_actual_s": {
                "count": 0,
                "max": null,
                "median": null
              }
            },
            {
              "arm_id": "direct-deepseek",
              "attempt_distribution": {
                "1": 50
              },
              "attempts": 50,
              "condition": "cold",
              "eventual_successes": 50,
              "exhausted": 0,
              "final_attempt_start_offset_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "first_attempt_successes": 50,
              "logical_cells": 50,
              "output_limits": {
                "4096": {
                  "equal": 0,
                  "measured": 50
                }
              },
              "recovery_completion_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "rescued": 0,
              "retried": 0,
              "retry_after_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "wait_actual_s": {
                "count": 0,
                "max": null,
                "median": null
              }
            },
            {
              "arm_id": "direct-deepseek",
              "attempt_distribution": {
                "1": 50
              },
              "attempts": 50,
              "condition": "warm",
              "eventual_successes": 50,
              "exhausted": 0,
              "final_attempt_start_offset_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "first_attempt_successes": 50,
              "logical_cells": 50,
              "output_limits": {
                "4096": {
                  "equal": 0,
                  "measured": 50
                }
              },
              "recovery_completion_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "rescued": 0,
              "retried": 0,
              "retry_after_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "wait_actual_s": {
                "count": 0,
                "max": null,
                "median": null
              }
            },
            {
              "arm_id": "openrouter-deepseek",
              "attempt_distribution": {
                "1": 50
              },
              "attempts": 50,
              "condition": "cold",
              "eventual_successes": 50,
              "exhausted": 0,
              "final_attempt_start_offset_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "first_attempt_successes": 50,
              "logical_cells": 50,
              "output_limits": {
                "4096": {
                  "equal": 0,
                  "measured": 50
                }
              },
              "recovery_completion_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "rescued": 0,
              "retried": 0,
              "retry_after_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "wait_actual_s": {
                "count": 0,
                "max": null,
                "median": null
              }
            },
            {
              "arm_id": "openrouter-deepseek",
              "attempt_distribution": {
                "1": 50
              },
              "attempts": 50,
              "condition": "warm",
              "eventual_successes": 50,
              "exhausted": 0,
              "final_attempt_start_offset_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "first_attempt_successes": 50,
              "logical_cells": 50,
              "output_limits": {
                "4096": {
                  "equal": 0,
                  "measured": 50
                }
              },
              "recovery_completion_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "rescued": 0,
              "retried": 0,
              "retry_after_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "wait_actual_s": {
                "count": 0,
                "max": null,
                "median": null
              }
            },
            {
              "arm_id": "vercel-deepseek",
              "attempt_distribution": {
                "1": 50
              },
              "attempts": 50,
              "condition": "cold",
              "eventual_successes": 50,
              "exhausted": 0,
              "final_attempt_start_offset_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "first_attempt_successes": 50,
              "logical_cells": 50,
              "output_limits": {
                "4096": {
                  "equal": 0,
                  "measured": 50
                }
              },
              "recovery_completion_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "rescued": 0,
              "retried": 0,
              "retry_after_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "wait_actual_s": {
                "count": 0,
                "max": null,
                "median": null
              }
            },
            {
              "arm_id": "vercel-deepseek",
              "attempt_distribution": {
                "1": 50
              },
              "attempts": 50,
              "condition": "warm",
              "eventual_successes": 50,
              "exhausted": 0,
              "final_attempt_start_offset_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "first_attempt_successes": 50,
              "logical_cells": 50,
              "output_limits": {
                "4096": {
                  "equal": 0,
                  "measured": 50
                }
              },
              "recovery_completion_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "rescued": 0,
              "retried": 0,
              "retry_after_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "wait_actual_s": {
                "count": 0,
                "max": null,
                "median": null
              }
            }
          ],
          "overall": {
            "attempt_distribution": {
              "1": 500
            },
            "attempts": 500,
            "eventual_successes": 500,
            "exhausted": 0,
            "final_attempt_start_offset_s": {
              "count": 0,
              "max": null,
              "median": null
            },
            "first_attempt_successes": 500,
            "logical_cells": 500,
            "output_limits": {
              "4096": {
                "equal": 0,
                "measured": 500
              }
            },
            "recovery_completion_s": {
              "count": 0,
              "max": null,
              "median": null
            },
            "rescued": 0,
            "retried": 0,
            "retry_after_s": {
              "count": 0,
              "max": null,
              "median": null
            },
            "wait_actual_s": {
              "count": 0,
              "max": null,
              "median": null
            }
          }
        },
        "run_note": "Five routes, each with 50 cold and 50 warm measured requests. Warm measurements include a separate primer. Gateway routes lock to DeepSeek with fallbacks and response caching disabled. Token counts are route-reported; Cloudflare reported a different input-token denominator from the other routes.",
        "schedule_digest": "95d76e5be01144a799299c29835dd8019985eb984841f87b588b2760e8a4972b",
        "scheduled_blocks_per_condition": 50,
        "submitter": "OpenBench",
        "title": "DeepSeek V4 Flash Gateway Bench, 100 requests per route",
        "track": "request_probe",
        "verified_with_commit": "e0e5dd2978c718c481926b5d293e874baea70eb1"
      },
      {
        "arms": [
          {
            "arm_id": "direct-moonshot",
            "baseline": true,
            "conditions": {
              "cold": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0025589999999999996,
                    "p95": 0.00417795
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 0,
                      "ratio": 0.0,
                      "total": 50
                    },
                    "p50": null,
                    "p95": null
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 115.0,
                    "p95": 117.55
                  },
                  "cold_end_to_end_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.3408743540057912,
                    "p95": 3.5371219464577734
                  },
                  "cold_end_to_end_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.3407649584987666,
                    "p95": 3.536984777424368
                  },
                  "cold_end_to_end_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.341465187491849,
                    "p95": 3.537503040069714
                  },
                  "cold_end_to_end_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 6.9630036875314545,
                    "p95": 12.424950442026601
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 115.0,
                    "p95": 117.55
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0025589999999999996,
                    "p95": 0.00417795
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0025589999999999996,
                    "p95": 0.00417795
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 147.5,
                    "p95": 255.34999999999994
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 6.932257208507508,
                    "p95": 12.390727415125001
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.256427521002479,
                    "p95": 3.493029477560776
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.256316937506199,
                    "p95": 3.492892308527371
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.2567228745028842,
                    "p95": 3.493410571172717
                  },
                  "setup_dns_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0019542289956007153,
                    "p95": 0.03742405395896639
                  },
                  "setup_tcp_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.011529354000231251,
                    "p95": 0.012927814101567492
                  },
                  "setup_tls_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.01823604150558822,
                    "p95": 0.021717573274509048
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 34.556510795925576,
                    "p95": 40.93176376357578
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 263.0,
                    "p95": 370.15
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              },
              "warm": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.004854,
                    "p95": 0.007339499999999997
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 0,
                      "ratio": 0.0,
                      "total": 50
                    },
                    "p50": null,
                    "p95": null
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 0.6,
                      "total": 50
                    },
                    "p50": 115.0,
                    "p95": 119.0
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 116.0,
                    "p95": 119.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.002424,
                    "p95": 0.004196849999999999
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.004854,
                    "p95": 0.007339499999999997
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 138.0,
                    "p95": 256.69999999999993
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 6.7980383959948085,
                    "p95": 11.103674260218394
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.350536354468204,
                    "p95": 3.7097707100241677
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.3503894789901096,
                    "p95": 3.709603760347818
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.3509444585070014,
                    "p95": 3.7102690287836593
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 34.24982392998798,
                    "p95": 56.64701735029892
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 256.0,
                    "p95": 372.15
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              }
            },
            "role": "direct"
          },
          {
            "arm_id": "cloudflare-moonshot",
            "baseline": false,
            "conditions": {
              "cold": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0020325,
                    "p95": 0.004336349999999999
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 0,
                      "ratio": 0.0,
                      "total": 50
                    },
                    "p50": null,
                    "p95": null
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 0.6,
                      "total": 50
                    },
                    "p50": 116.0,
                    "p95": 117.55
                  },
                  "cold_end_to_end_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.861416999978246,
                    "p95": 5.284222987203974
                  },
                  "cold_end_to_end_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.8613371459941845,
                    "p95": 5.284060535297611
                  },
                  "cold_end_to_end_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.8614580829744227,
                    "p95": 5.284300162189174
                  },
                  "cold_end_to_end_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 6.840523854014464,
                    "p95": 13.278519981409772
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 115.0,
                    "p95": 117.55
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0020325,
                    "p95": 0.004336349999999999
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0020325,
                    "p95": 0.004336349999999999
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 112.5,
                    "p95": 266.44999999999993
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 6.78791804148932,
                    "p95": 13.250897272530711
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.8314286670065485,
                    "p95": 5.25607548111584
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.831348813022487,
                    "p95": 5.255913029209477
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.831469750002725,
                    "p95": 5.2561526561010385
                  },
                  "setup_dns_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0015685624675825238,
                    "p95": 0.02148034388956148
                  },
                  "setup_tcp_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.01147183298598975,
                    "p95": 0.012824224834912457
                  },
                  "setup_tls_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.01623614551499486,
                    "p95": 0.019968068579328244
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 34.780247510437356,
                    "p95": 47.98312845429244
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 227.5,
                    "p95": 379.65
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              },
              "warm": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0050355,
                    "p95": 0.0070656
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 0,
                      "ratio": 0.0,
                      "total": 50
                    },
                    "p50": null,
                    "p95": null
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 42,
                      "ratio": 0.84,
                      "total": 50
                    },
                    "p50": 116.0,
                    "p95": 119.0
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 116.0,
                    "p95": 119.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.002157,
                    "p95": 0.0037338
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0050355,
                    "p95": 0.0070656
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 120.5,
                    "p95": 225.64999999999998
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 5.965509479487082,
                    "p95": 9.987062258317017
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.34417262498755,
                    "p95": 4.020355535333509
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.3439068954903632,
                    "p95": 4.020078435036702
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.344265291496413,
                    "p95": 4.020493122833431
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 34.913792524993696,
                    "p95": 43.99274982798177
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 237.0,
                    "p95": 342.0
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              }
            },
            "role": "gateway"
          },
          {
            "arm_id": "concentrate-moonshot",
            "baseline": false,
            "conditions": {
              "cold": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.001974,
                    "p95": 0.00384615
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 114.0,
                    "p95": 117.0
                  },
                  "cold_end_to_end_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.7304988539835904,
                    "p95": 3.893284224980744
                  },
                  "cold_end_to_end_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.7303663335042074,
                    "p95": 3.8932014001155038
                  },
                  "cold_end_to_end_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.7304988539835904,
                    "p95": 3.893284224980744
                  },
                  "cold_end_to_end_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 5.736859499505954,
                    "p95": 10.754648496073788
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 115.0,
                    "p95": 117.55
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.001974,
                    "p95": 0.00384615
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.001974,
                    "p95": 0.00384615
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 108.5,
                    "p95": 232.74999999999997
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 5.706972437008517,
                    "p95": 10.722715008535307
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.699270395503845,
                    "p95": 3.839293262208229
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.699137875024462,
                    "p95": 3.83921043734299
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.699270395503845,
                    "p95": 3.839293262208229
                  },
                  "setup_dns_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0014466459979303181,
                    "p95": 0.05600984998745842
                  },
                  "setup_tcp_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.011491000012028962,
                    "p95": 0.012474912649486213
                  },
                  "setup_tls_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0176489369769115,
                    "p95": 0.020750364658306354
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 36.20658051347921,
                    "p95": 45.27462216917274
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 224.0,
                    "p95": 351.04999999999995
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              },
              "warm": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0043485,
                    "p95": 0.0066783
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 115.0,
                    "p95": 118.0
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 116.0,
                    "p95": 119.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0023610000000000003,
                    "p95": 0.003992699999999999
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0043485,
                    "p95": 0.0066783
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 134.0,
                    "p95": 243.24999999999994
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 6.67688958303188,
                    "p95": 10.043060175035498
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.5143570005020592,
                    "p95": 3.6910123665293213
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.513396646012552,
                    "p95": 3.6908935290761296
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.5143570005020592,
                    "p95": 3.6910123665293213
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 34.90726580243485,
                    "p95": 44.36400666062325
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 251.5,
                    "p95": 357.9
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              }
            },
            "role": "gateway"
          },
          {
            "arm_id": "openrouter-moonshot",
            "baseline": false,
            "conditions": {
              "cold": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.002169,
                    "p95": 0.0039082499999999985
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 117.0
                  },
                  "cold_end_to_end_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.5931841249985155,
                    "p95": 3.415128739571082
                  },
                  "cold_end_to_end_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.5930756664893124,
                    "p95": 3.4150134478404643
                  },
                  "cold_end_to_end_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.6309518960188143,
                    "p95": 3.4294422376638973
                  },
                  "cold_end_to_end_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 6.040993520990014,
                    "p95": 10.189295322235555
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 115.0,
                    "p95": 117.55
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.002169,
                    "p95": 0.0039082499999999985
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.002169,
                    "p95": 0.0039082499999999985
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 121.5,
                    "p95": 237.34999999999997
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 6.011471833509859,
                    "p95": 10.151878664398097
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.537568145489786,
                    "p95": 3.3772392811020833
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.537497749493923,
                    "p95": 3.3771239893714657
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.5683611670101527,
                    "p95": 3.3915527791948983
                  },
                  "setup_dns_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0014995415403973311,
                    "p95": 0.02194011683168355
                  },
                  "setup_tcp_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.01107883348595351,
                    "p95": 0.012369329636567273
                  },
                  "setup_tls_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.015847895527258515,
                    "p95": 0.018659743608441203
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 35.117944885870706,
                    "p95": 47.96639300243133
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 237.0,
                    "p95": 353.34999999999997
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              },
              "warm": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.00489,
                    "p95": 0.0072232500000000005
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0,
                    "p95": 118.0
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 116.0,
                    "p95": 119.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.002517,
                    "p95": 0.0038145
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.00489,
                    "p95": 0.0072232500000000005
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 144.5,
                    "p95": 231.1
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 6.579976895503933,
                    "p95": 9.580938294032238
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.4234466455236543,
                    "p95": 2.982058181258617
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.423233875015285,
                    "p95": 2.9815335019200573
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.4363150210119784,
                    "p95": 2.9881874584709296
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 34.0894979982138,
                    "p95": 41.12953775944523
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 260.5,
                    "p95": 347.1
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              }
            },
            "role": "gateway"
          },
          {
            "arm_id": "vercel-moonshot",
            "baseline": false,
            "conditions": {
              "cold": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.002316,
                    "p95": 0.0039037499999999992
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 0,
                      "ratio": 0.0,
                      "total": 50
                    },
                    "p50": null,
                    "p95": null
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 113.0,
                    "p95": 117.55
                  },
                  "cold_end_to_end_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.7901442915317602,
                    "p95": 4.549596450017996
                  },
                  "cold_end_to_end_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.781087291514268,
                    "p95": 4.540028418789733
                  },
                  "cold_end_to_end_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.7901442915317602,
                    "p95": 4.549596450017996
                  },
                  "cold_end_to_end_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 6.574096833006479,
                    "p95": 11.279137716171679
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 115.0,
                    "p95": 117.55
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.002316,
                    "p95": 0.0039037499999999992
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.002316,
                    "p95": 0.0039037499999999992
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 131.5,
                    "p95": 237.39999999999998
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 6.541199625004083,
                    "p95": 11.248750870797084
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.7338413545221556,
                    "p95": 4.5036803084280095
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.725528958515497,
                    "p95": 4.494112277199745
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.7338413545221556,
                    "p95": 4.5036803084280095
                  },
                  "setup_dns_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.0019090414862148464,
                    "p95": 0.13546928298310323
                  },
                  "setup_tcp_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.005395728978328407,
                    "p95": 0.006585891672875732
                  },
                  "setup_tls_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.023584353970363736,
                    "p95": 0.05281281424104223
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 34.70185523001546,
                    "p95": 48.12506506678979
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 244.5,
                    "p95": 351.65
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              },
              "warm": {
                "availability": {
                  "attempted": 50,
                  "rate": 1.0,
                  "successes": 50,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.9286499658256813
                  }
                },
                "denominators": {
                  "attempted": 50,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 50,
                  "scheduled": 50,
                  "success": 50
                },
                "metrics": {
                  "budget_debit_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.004595999999999999,
                    "p95": 0.007244699999999999
                  },
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 0,
                      "ratio": 0.0,
                      "total": 50
                    },
                    "p50": null,
                    "p95": null
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 115.0,
                    "p95": 119.0
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 116.0,
                    "p95": 119.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.002388,
                    "p95": 0.0040272
                  },
                  "observed_cost_usd": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 0.004595999999999999,
                    "p95": 0.007244699999999999
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 136.0,
                    "p95": 245.29999999999998
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 6.319876166497124,
                    "p95": 10.412645818750025
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.4590459165046923,
                    "p95": 3.1380958689202085
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.4588287915103137,
                    "p95": 3.1321478919911896
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 2.4590459165046923,
                    "p95": 3.1380958689202085
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 35.13270986441798,
                    "p95": 45.9708967969119
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 50,
                      "ratio": 1.0,
                      "total": 50
                    },
                    "p50": 252.0,
                    "p95": 361.2
                  }
                },
                "retry_diagnostics": {
                  "attempt_count_distribution": {
                    "1": 50
                  },
                  "eventual_successes": 50,
                  "first_attempt_successes": 50,
                  "logical_cells": 50,
                  "retry_rescues": 0,
                  "unknown_cost_attempts": 0
                }
              }
            },
            "role": "gateway"
          }
        ],
        "baseline_arm_id": "direct-moonshot",
        "complete_blocks": {
          "cold": 50,
          "warm": 50
        },
        "completion_integrity": {
          "arms": {
            "cloudflare-moonshot": {
              "cold": {
                "measured": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                }
              },
              "warm": {
                "measured": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                },
                "warm_primer": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                }
              }
            },
            "concentrate-moonshot": {
              "cold": {
                "measured": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                }
              },
              "warm": {
                "measured": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                },
                "warm_primer": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                }
              }
            },
            "direct-moonshot": {
              "cold": {
                "measured": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                }
              },
              "warm": {
                "measured": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                },
                "warm_primer": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                }
              }
            },
            "openrouter-moonshot": {
              "cold": {
                "measured": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                }
              },
              "warm": {
                "measured": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                },
                "warm_primer": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                }
              }
            },
            "vercel-moonshot": {
              "cold": {
                "measured": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                }
              },
              "warm": {
                "measured": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                },
                "warm_primer": {
                  "length": 0,
                  "missing": 0,
                  "natural_stop": 50,
                  "other": 0,
                  "total": 50
                }
              }
            }
          }
        },
        "contrasts": [
          {
            "arm_id": "cloudflare-moonshot",
            "conditions": {
              "cold": {
                "budget_debit_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00027000000000000006,
                    "low": -0.0007426875000000001,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -9.750000000000026e-05
                },
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 0,
                    "ratio": 0.0,
                    "total": 50
                  },
                  "interval": null,
                  "median_gateway_minus_direct": null
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 0.6,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cold_end_to_end_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.6822057920508087,
                    "low": 0.24033536751085194,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.40836185446823947
                },
                "cold_end_to_end_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.6821175830555148,
                    "low": 0.23213568751816638,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.40837931248825043
                },
                "cold_end_to_end_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.6849017289932817,
                    "low": 0.24036879098275676,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.4081882709870115
                },
                "cold_end_to_end_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.3800547230959634,
                    "low": -1.4920864169835113,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.18733412501751445
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00027000000000000006,
                    "low": -0.0007576875,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -9.750000000000026e-05
                },
                "observed_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00027000000000000006,
                    "low": -0.0007650000000000001,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -9.750000000000026e-05
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 18.0,
                    "low": -50.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -6.5
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.4276820624945685,
                    "low": -1.4993268134712707,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.1836243955185637
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.6739823125244584,
                    "low": 0.22048431300208904,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.41258220802410506
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.6738972285529599,
                    "low": 0.22179816802963614,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.412599666044116
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.6739592500089202,
                    "low": 0.22217862505931407,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.4124086245428771
                },
                "setup_dns_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -8.470853208564222e-05,
                    "low": -0.00044774997513741255,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.00019064504886046052
                },
                "setup_tcp_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00035220797872170806,
                    "low": -0.00035712544922716916,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -5.841749953106046e-05
                },
                "setup_tls_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -0.0011573954834602773,
                    "low": -0.00343504105694592,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.0025443755148444325
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 3.8817084575759253,
                    "low": -0.4727888248150727,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 1.0745203233789198
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 18.0,
                    "low": -50.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -6.5
                }
              },
              "warm": {
                "budget_debit_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0002474999999999999,
                    "low": -0.0006149999999999999,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -7.500000000000041e-05
                },
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 0,
                    "ratio": 0.0,
                    "total": 50
                  },
                  "interval": null,
                  "median_gateway_minus_direct": null
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 22,
                    "ratio": 0.44,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00011999999999999988,
                    "low": -0.0007950000000000001,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.00018000000000000004
                },
                "observed_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00023999999999999976,
                    "low": -0.0005197499999999998,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -7.500000000000041e-05
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 8.0,
                    "low": -53.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -12.0
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.2165948379952123,
                    "low": -2.418671292020008,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.3198478120029904
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.14833699065784445,
                    "low": -0.06755643748329021,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.06709331250749528
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.15941572899464518,
                    "low": -0.06759045898797922,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.06708529201569036
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.1591755619738251,
                    "low": -0.0677384169830475,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.06692560450755991
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 4.39026070271299,
                    "low": -1.4148920903602829,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 2.198567063389657
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 9.0,
                    "low": -57.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -12.0
                }
              }
            },
            "direct_arm": "direct-moonshot"
          },
          {
            "arm_id": "concentrate-moonshot",
            "conditions": {
              "cold": {
                "budget_debit_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00014249999999999994,
                    "low": -0.0009749999999999997,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.00019500000000000008
                },
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 0,
                    "ratio": 0.0,
                    "total": 50
                  },
                  "interval": null,
                  "median_gateway_minus_direct": null
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cold_end_to_end_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.3930217661647472,
                    "low": 0.05500431201653555,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.20917179150274023
                },
                "cold_end_to_end_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.3929418759653345,
                    "low": 0.07266206331405559,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.20910174999153242
                },
                "cold_end_to_end_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.3868742997146907,
                    "low": 0.07820137502858415,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.20894820801913738
                },
                "cold_end_to_end_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.5973289579560515,
                    "low": -2.051774082996417,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.02572331254486926
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00013499999999999992,
                    "low": -0.0009824999999999999,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.00019500000000000008
                },
                "observed_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00014249999999999994,
                    "low": -0.0009824999999999999,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.00019500000000000008
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 9.5,
                    "low": -64.0125,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -13.0
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.664924874989083,
                    "low": -2.0705626669805497,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.024639271025080234
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.4103910851081286,
                    "low": 0.08604708302300423,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.25081060451339
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.39302370900986716,
                    "low": 0.0860375000629574,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.24913983399164863
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.3926241250592284,
                    "low": 0.07736402048612945,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.25061056250706315
                },
                "setup_dns_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -0.00011385398101992905,
                    "low": -0.0005720000481233001,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.0002643335028551519
                },
                "setup_tcp_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00019458401948213577,
                    "low": -0.000433248991612345,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.00024377048248425126
                },
                "setup_tls_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 5.339551717042923e-05,
                    "low": -0.0015801669796928763,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.0004316250269766897
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 2.510007077664575,
                    "low": -0.47301897680970306,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 1.1016940481901756
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 9.0,
                    "low": -66.5,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -13.0
                }
              },
              "warm": {
                "budget_debit_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00012000000000000031,
                    "low": -0.0008999999999999998,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.00030750000000000005
                },
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 0,
                    "ratio": 0.0,
                    "total": 50
                  },
                  "interval": null,
                  "median_gateway_minus_direct": null
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 0.6,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": -112.5,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00069,
                    "low": -0.0005550000000000002,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0001200000000000001
                },
                "observed_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0001275,
                    "low": -0.0009000000000000002,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.00030750000000000005
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 46.0,
                    "low": -35.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 8.0
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.543621791061014,
                    "low": -1.2512867914920207,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.39320156202302314
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.28082520805764943,
                    "low": 0.11046170850750059,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.2362370619957801
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.28074600000400096,
                    "low": 0.11042906303191558,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.2362701450183522
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.30280241649597883,
                    "low": 0.11038214602740481,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.23603631200967357
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 2.3157978002251554,
                    "low": -3.402426369001674,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.488899496685395
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 46.0,
                    "low": -35.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 8.0
                }
              }
            },
            "direct_arm": "direct-moonshot"
          },
          {
            "arm_id": "openrouter-moonshot",
            "conditions": {
              "cold": {
                "budget_debit_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00048000000000000007,
                    "low": -0.0006300000000000004,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.0004124999999999999
                },
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 0,
                    "ratio": 0.0,
                    "total": 50
                  },
                  "interval": null,
                  "median_gateway_minus_direct": null
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -113.0,
                    "low": -115.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -114.0
                },
                "cold_end_to_end_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.41198291699402034,
                    "low": 0.08016537496587262,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.23337768798228353
                },
                "cold_end_to_end_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.44914772899937816,
                    "low": 0.07672368752537295,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.23340031248517334
                },
                "cold_end_to_end_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.42750266697839834,
                    "low": 0.09169000101974234,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.24215775047196075
                },
                "cold_end_to_end_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.5509323329897597,
                    "low": -1.6535358329419978,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.4945721665280871
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0004350000000000003,
                    "low": -0.0006300000000000004,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.0004124999999999999
                },
                "observed_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0004350000000000003,
                    "low": -0.0006300000000000004,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.0004124999999999999
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 32.0,
                    "low": -42.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -27.5
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.5767253539524972,
                    "low": -1.6975914380163886,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.5094129790086299
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.40981477044988424,
                    "low": 0.07079074997454882,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.25863808300346136
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.40853566699661314,
                    "low": 0.07059245886266581,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.2586524369835388
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.4201126239495352,
                    "low": 0.09001716700731777,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.27101972899981774
                },
                "setup_dns_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -0.0001827677100664031,
                    "low": -0.0007047080434858799,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.0003369790210854262
                },
                "setup_tcp_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -1.010397681966424e-05,
                    "low": -0.0005656879802700132,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.0003016249684151262
                },
                "setup_tls_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -0.0017642920138314366,
                    "low": -0.003255791962146759,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.0023695209820289165
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 4.919683181606988,
                    "low": -1.1588766092378755,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 1.4697893945012979
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 29.0,
                    "low": -42.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -27.5
                }
              },
              "warm": {
                "budget_debit_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0007049999999999999,
                    "low": -0.0005025000000000003,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.00014999999999999996
                },
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 0,
                    "ratio": 0.0,
                    "total": 50
                  },
                  "interval": null,
                  "median_gateway_minus_direct": null
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 0.6,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -114.5,
                    "low": -116.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -115.0
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0005101874999999988,
                    "low": -0.0005324999999999999,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.00013500000000000003
                },
                "observed_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0005999999999999998,
                    "low": -0.0005025000000000003,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.00014999999999999996
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 39.0,
                    "low": -35.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -9.0
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.6201724994753022,
                    "low": -1.5524240009835921,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.30295506250695325
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.37810760451247916,
                    "low": -0.07151372896623798,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.18709006253629923
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.3471401250571944,
                    "low": -0.06509566598106176,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.18714883300708607
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.37796237503062,
                    "low": -0.07737935348995961,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.192866645520553
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 2.3853801632398755,
                    "low": -2.546253124610786,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.44630866589768914
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 33.0,
                    "low": -35.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -9.0
                }
              }
            },
            "direct_arm": "direct-moonshot"
          },
          {
            "arm_id": "vercel-moonshot",
            "conditions": {
              "cold": {
                "budget_debit_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": -0.00043500000000000006,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.00023999999999999998
                },
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 0,
                    "ratio": 0.0,
                    "total": 50
                  },
                  "interval": null,
                  "median_gateway_minus_direct": null
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": -113.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cold_end_to_end_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.4829652914777398,
                    "low": 0.11643183400155976,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.30618385353591293
                },
                "cold_end_to_end_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.46763537498191,
                    "low": 0.11323948000790551,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.29700564546510577
                },
                "cold_end_to_end_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.47388991696061566,
                    "low": 0.11603683396242559,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.30595893700956367
                },
                "cold_end_to_end_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.1290579589549452,
                    "low": -1.1103186664986424,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.1286372504546307
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": -0.00043499999999999995,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.00023999999999999998
                },
                "observed_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.500000000000004e-05,
                    "low": -0.00043500000000000006,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.00023999999999999998
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": -29.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -16.0
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.1291821660124697,
                    "low": -1.0472130619746167,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.133951646508649
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.49460370794986375,
                    "low": 0.12750541599234566,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.30621181303285994
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.4922694780107122,
                    "low": 0.11826262500835583,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.29701895799371414
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.4981667494866997,
                    "low": 0.12730983295477927,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.30598366699996404
                },
                "setup_dns_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0004861254710704088,
                    "low": -0.0004438891977770254,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 5.231300019659102e-05
                },
                "setup_tcp_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -0.005884960866387706,
                    "low": -0.006372882567666238,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.006230812956346199
                },
                "setup_tls_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.005567874992266297,
                    "low": 0.004549832985503599,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0050677919934969395
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 2.433766111050348,
                    "low": 0.5777443242864901,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 1.7278861760896582
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": -29.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -16.0
                }
              },
              "warm": {
                "budget_debit_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0005174999999999997,
                    "low": -0.00033000000000000043,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 5.2500000000000246e-05
                },
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 0,
                    "ratio": 0.0,
                    "total": 50
                  },
                  "interval": null,
                  "median_gateway_minus_direct": null
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 0.6,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0003225000000000002,
                    "low": -0.0005999999999999998,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.00010500000000000006
                },
                "observed_cost_usd": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.000502874999999998,
                    "low": -0.0003375000000000001,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 5.2500000000000246e-05
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 20.5,
                    "low": -40.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -7.0
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.1453644995053764,
                    "low": -0.8916850004752632,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.3357554165122565
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.2823909370054025,
                    "low": -0.05229558298015036,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.12231924996012822
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.2712806239724159,
                    "low": -0.05553722047916381,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.11725687448051758
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.2819517284806352,
                    "low": -0.060310375498374924,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.12223410396836698
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 2.6568343795169227,
                    "low": -1.6771585061636287,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.4888619318136662
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 50,
                    "ratio": 1.0,
                    "total": 50
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 21.0,
                    "low": -40.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -7.0
                }
              }
            },
            "direct_arm": "direct-moonshot"
          }
        ],
        "date": "2026-07-28",
        "default": false,
        "experiment_digest": "9748aef6256df8feeb4f0a075f7cfce5e009aad22d029c55795e5af354550f6c",
        "experiment_id": "kimi-k3-five-way-chat-r100-4096-retry-v1-2026-07-28",
        "experiment_path": "gateway-probe/2026-07-28-kimi-k3-managed-100/experiment.json",
        "family": "gateway_probe",
        "id": "2026-07-28-kimi-k3-managed-100",
        "kind": "release",
        "link": null,
        "model": "Kimi K3",
        "model_match": "rolling_alias",
        "output_token_limit_equalities": {
          "arms": {
            "cloudflare-moonshot": {
              "cold": {
                "equal": 0,
                "measured": 50
              },
              "warm": {
                "equal": 0,
                "measured": 50
              }
            },
            "concentrate-moonshot": {
              "cold": {
                "equal": 0,
                "measured": 50
              },
              "warm": {
                "equal": 0,
                "measured": 50
              }
            },
            "direct-moonshot": {
              "cold": {
                "equal": 0,
                "measured": 50
              },
              "warm": {
                "equal": 0,
                "measured": 50
              }
            },
            "openrouter-moonshot": {
              "cold": {
                "equal": 0,
                "measured": 50
              },
              "warm": {
                "equal": 0,
                "measured": 50
              }
            },
            "vercel-moonshot": {
              "cold": {
                "equal": 0,
                "measured": 50
              },
              "warm": {
                "equal": 0,
                "measured": 50
              }
            }
          },
          "configured_limit": 4096
        },
        "path": null,
        "price_digest": "cc94c4d7f77608a63623d40a620256840ea09247c4a114dd55a61a8ff5caf0a7",
        "result_count": 500,
        "result_schema_version": 4,
        "results_path": "gateway-probe/2026-07-28-kimi-k3-managed-100/results.jsonl",
        "retry_summary": {
          "groups": [
            {
              "arm_id": "cloudflare-moonshot",
              "attempt_distribution": {
                "1": 50
              },
              "attempts": 50,
              "condition": "cold",
              "eventual_successes": 50,
              "exhausted": 0,
              "final_attempt_start_offset_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "first_attempt_successes": 50,
              "logical_cells": 50,
              "output_limits": {
                "4096": {
                  "equal": 0,
                  "measured": 50
                }
              },
              "recovery_completion_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "rescued": 0,
              "retried": 0,
              "retry_after_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "wait_actual_s": {
                "count": 0,
                "max": null,
                "median": null
              }
            },
            {
              "arm_id": "cloudflare-moonshot",
              "attempt_distribution": {
                "1": 50
              },
              "attempts": 50,
              "condition": "warm",
              "eventual_successes": 50,
              "exhausted": 0,
              "final_attempt_start_offset_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "first_attempt_successes": 50,
              "logical_cells": 50,
              "output_limits": {
                "4096": {
                  "equal": 0,
                  "measured": 50
                }
              },
              "recovery_completion_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "rescued": 0,
              "retried": 0,
              "retry_after_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "wait_actual_s": {
                "count": 0,
                "max": null,
                "median": null
              }
            },
            {
              "arm_id": "concentrate-moonshot",
              "attempt_distribution": {
                "1": 50
              },
              "attempts": 50,
              "condition": "cold",
              "eventual_successes": 50,
              "exhausted": 0,
              "final_attempt_start_offset_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "first_attempt_successes": 50,
              "logical_cells": 50,
              "output_limits": {
                "4096": {
                  "equal": 0,
                  "measured": 50
                }
              },
              "recovery_completion_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "rescued": 0,
              "retried": 0,
              "retry_after_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "wait_actual_s": {
                "count": 0,
                "max": null,
                "median": null
              }
            },
            {
              "arm_id": "concentrate-moonshot",
              "attempt_distribution": {
                "1": 50
              },
              "attempts": 50,
              "condition": "warm",
              "eventual_successes": 50,
              "exhausted": 0,
              "final_attempt_start_offset_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "first_attempt_successes": 50,
              "logical_cells": 50,
              "output_limits": {
                "4096": {
                  "equal": 0,
                  "measured": 50
                }
              },
              "recovery_completion_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "rescued": 0,
              "retried": 0,
              "retry_after_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "wait_actual_s": {
                "count": 0,
                "max": null,
                "median": null
              }
            },
            {
              "arm_id": "direct-moonshot",
              "attempt_distribution": {
                "1": 50
              },
              "attempts": 50,
              "condition": "cold",
              "eventual_successes": 50,
              "exhausted": 0,
              "final_attempt_start_offset_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "first_attempt_successes": 50,
              "logical_cells": 50,
              "output_limits": {
                "4096": {
                  "equal": 0,
                  "measured": 50
                }
              },
              "recovery_completion_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "rescued": 0,
              "retried": 0,
              "retry_after_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "wait_actual_s": {
                "count": 0,
                "max": null,
                "median": null
              }
            },
            {
              "arm_id": "direct-moonshot",
              "attempt_distribution": {
                "1": 50
              },
              "attempts": 50,
              "condition": "warm",
              "eventual_successes": 50,
              "exhausted": 0,
              "final_attempt_start_offset_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "first_attempt_successes": 50,
              "logical_cells": 50,
              "output_limits": {
                "4096": {
                  "equal": 0,
                  "measured": 50
                }
              },
              "recovery_completion_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "rescued": 0,
              "retried": 0,
              "retry_after_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "wait_actual_s": {
                "count": 0,
                "max": null,
                "median": null
              }
            },
            {
              "arm_id": "openrouter-moonshot",
              "attempt_distribution": {
                "1": 50
              },
              "attempts": 50,
              "condition": "cold",
              "eventual_successes": 50,
              "exhausted": 0,
              "final_attempt_start_offset_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "first_attempt_successes": 50,
              "logical_cells": 50,
              "output_limits": {
                "4096": {
                  "equal": 0,
                  "measured": 50
                }
              },
              "recovery_completion_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "rescued": 0,
              "retried": 0,
              "retry_after_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "wait_actual_s": {
                "count": 0,
                "max": null,
                "median": null
              }
            },
            {
              "arm_id": "openrouter-moonshot",
              "attempt_distribution": {
                "1": 50
              },
              "attempts": 50,
              "condition": "warm",
              "eventual_successes": 50,
              "exhausted": 0,
              "final_attempt_start_offset_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "first_attempt_successes": 50,
              "logical_cells": 50,
              "output_limits": {
                "4096": {
                  "equal": 0,
                  "measured": 50
                }
              },
              "recovery_completion_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "rescued": 0,
              "retried": 0,
              "retry_after_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "wait_actual_s": {
                "count": 0,
                "max": null,
                "median": null
              }
            },
            {
              "arm_id": "vercel-moonshot",
              "attempt_distribution": {
                "1": 50
              },
              "attempts": 50,
              "condition": "cold",
              "eventual_successes": 50,
              "exhausted": 0,
              "final_attempt_start_offset_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "first_attempt_successes": 50,
              "logical_cells": 50,
              "output_limits": {
                "4096": {
                  "equal": 0,
                  "measured": 50
                }
              },
              "recovery_completion_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "rescued": 0,
              "retried": 0,
              "retry_after_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "wait_actual_s": {
                "count": 0,
                "max": null,
                "median": null
              }
            },
            {
              "arm_id": "vercel-moonshot",
              "attempt_distribution": {
                "1": 50
              },
              "attempts": 50,
              "condition": "warm",
              "eventual_successes": 50,
              "exhausted": 0,
              "final_attempt_start_offset_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "first_attempt_successes": 50,
              "logical_cells": 50,
              "output_limits": {
                "4096": {
                  "equal": 0,
                  "measured": 50
                }
              },
              "recovery_completion_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "rescued": 0,
              "retried": 0,
              "retry_after_s": {
                "count": 0,
                "max": null,
                "median": null
              },
              "wait_actual_s": {
                "count": 0,
                "max": null,
                "median": null
              }
            }
          ],
          "overall": {
            "attempt_distribution": {
              "1": 500
            },
            "attempts": 500,
            "eventual_successes": 500,
            "exhausted": 0,
            "final_attempt_start_offset_s": {
              "count": 0,
              "max": null,
              "median": null
            },
            "first_attempt_successes": 500,
            "logical_cells": 500,
            "output_limits": {
              "4096": {
                "equal": 0,
                "measured": 500
              }
            },
            "recovery_completion_s": {
              "count": 0,
              "max": null,
              "median": null
            },
            "rescued": 0,
            "retried": 0,
            "retry_after_s": {
              "count": 0,
              "max": null,
              "median": null
            },
            "wait_actual_s": {
              "count": 0,
              "max": null,
              "median": null
            }
          }
        },
        "run_note": "Five routes, each with 50 cold and 50 warm measured requests. Warm measurements include a separate primer. Routes use rolling Kimi K3 aliases.",
        "schedule_digest": "07d421331cd2a61539e7cb4cb3d3ff5de3dfc2ed16cc91d85292e71efd60b01c",
        "scheduled_blocks_per_condition": 50,
        "submitter": "OpenBench",
        "title": "Kimi K3 Gateway Bench, 100 requests per route",
        "track": "request_probe",
        "verified_with_commit": "c45441997c8181ddd7af865e2726f8548296790b"
      },
      {
        "arms": [
          {
            "arm_id": "direct-openai",
            "baseline": true,
            "conditions": {
              "cold": {
                "availability": {
                  "attempted": 30,
                  "rate": 1.0,
                  "successes": 30,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.886482908609522
                  }
                },
                "denominators": {
                  "attempted": 30,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 30,
                  "scheduled": 30,
                  "success": 30
                },
                "metrics": {
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "charged_cost_usd": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 1.395e-05,
                    "p95": 1.4399999999999998e-05
                  },
                  "cold_end_to_end_first_body_byte_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.2618867295095697,
                    "p95": 0.9161389394081192
                  },
                  "cold_end_to_end_response_headers_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.26175627097836696,
                    "p95": 0.9146265079878497
                  },
                  "cold_end_to_end_semantic_ttft_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.5556630629871506,
                    "p95": 1.384148770500905
                  },
                  "cold_end_to_end_stream_total_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.7653896670381073,
                    "p95": 1.7911753040534668
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 37.0,
                    "p95": 39.55
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 1.395e-05,
                    "p95": 1.4399999999999998e-05
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 14.0,
                    "p95": 14.0
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.7349729375273455,
                    "p95": 1.7633883562404655
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.2320579580264166,
                    "p95": 0.8883519915951177
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.2319540415192023,
                    "p95": 0.8868395601748482
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.5268897500063758,
                    "p95": 1.3563618226879037
                  },
                  "setup_dns_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.0016002080228645355,
                    "p95": 0.002891178955906066
                  },
                  "setup_tcp_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.011436458502430469,
                    "p95": 0.012133112322771922
                  },
                  "setup_tls_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.01643262500874698,
                    "p95": 0.022621695898124017
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 92.29271620360342,
                    "p95": 123.19624845830998
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 51.0,
                    "p95": 54.0
                  }
                }
              },
              "warm": {
                "availability": {
                  "attempted": 30,
                  "rate": 1.0,
                  "successes": 30,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.886482908609522
                  }
                },
                "denominators": {
                  "attempted": 30,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 30,
                  "scheduled": 30,
                  "success": 30
                },
                "metrics": {
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "charged_cost_usd": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 2.835e-05,
                    "p95": 3.0300000000000005e-05
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 38.0,
                    "p95": 40.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 1.41e-05,
                    "p95": 1.5457499999999996e-05
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 14.0,
                    "p95": 16.19999999999999
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.6562053334782831,
                    "p95": 1.2209905081253964
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.18428431247593835,
                    "p95": 0.43743395827186743
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.1841714994807262,
                    "p95": 0.4361857331299687
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.5037482295010705,
                    "p95": 0.9772484541055746
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 95.10098558742803,
                    "p95": 127.21719451628391
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 52.0,
                    "p95": 54.55
                  }
                }
              }
            },
            "role": "direct"
          },
          {
            "arm_id": "cloudflare-openai",
            "baseline": false,
            "conditions": {
              "cold": {
                "availability": {
                  "attempted": 30,
                  "rate": 1.0,
                  "successes": 30,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.886482908609522
                  }
                },
                "denominators": {
                  "attempted": 30,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 30,
                  "scheduled": 30,
                  "success": 30
                },
                "metrics": {
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "charged_cost_usd": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 1.395e-05,
                    "p95": 1.4332499999999999e-05
                  },
                  "cold_end_to_end_first_body_byte_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.3778885630017612,
                    "p95": 0.8001176937425039
                  },
                  "cold_end_to_end_response_headers_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.37778033351060003,
                    "p95": 0.7999479419959239
                  },
                  "cold_end_to_end_semantic_ttft_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.684436937503051,
                    "p95": 1.0906790498644112
                  },
                  "cold_end_to_end_stream_total_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.8488696874992456,
                    "p95": 1.247032909898553
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 37.0,
                    "p95": 39.55
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 1.395e-05,
                    "p95": 1.4332499999999999e-05
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 14.0,
                    "p95": 14.0
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.8121094164962415,
                    "p95": 1.2154791437787935
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.34092250050161965,
                    "p95": 0.7672035979398056
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.34081427101045847,
                    "p95": 0.7670338461932256
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.6555284794885665,
                    "p95": 1.054727060740697
                  },
                  "setup_dns_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.0013577920035459101,
                    "p95": 0.0017405648017302154
                  },
                  "setup_tcp_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.01132097901427187,
                    "p95": 0.0123495559731964
                  },
                  "setup_tls_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.016649396013235673,
                    "p95": 0.0229986027610721
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 84.58266220944367,
                    "p95": 113.73809479723356
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 51.0,
                    "p95": 53.55
                  }
                }
              },
              "warm": {
                "availability": {
                  "attempted": 30,
                  "rate": 1.0,
                  "successes": 30,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.886482908609522
                  }
                },
                "denominators": {
                  "attempted": 30,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 30,
                  "scheduled": 30,
                  "success": 30
                },
                "metrics": {
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "charged_cost_usd": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 2.835e-05,
                    "p95": 3.0300000000000005e-05
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 38.0,
                    "p95": 40.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 1.41e-05,
                    "p95": 1.5457499999999996e-05
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 14.0,
                    "p95": 16.19999999999999
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.7536033329670317,
                    "p95": 0.8933024374855449
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.26018185398424976,
                    "p95": 0.4104486996948252
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.2600222500041127,
                    "p95": 0.41031452936585977
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.5606690415006597,
                    "p95": 0.739080439400277
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 96.57889187473157,
                    "p95": 126.43455910330448
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 52.0,
                    "p95": 54.55
                  }
                }
              }
            },
            "role": "gateway"
          },
          {
            "arm_id": "concentrate-openai",
            "baseline": false,
            "conditions": {
              "cold": {
                "availability": {
                  "attempted": 30,
                  "rate": 1.0,
                  "successes": 30,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.886482908609522
                  }
                },
                "denominators": {
                  "attempted": 30,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 30,
                  "scheduled": 30,
                  "success": 30
                },
                "metrics": {
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "charged_cost_usd": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 1.395e-05,
                    "p95": 1.4399999999999998e-05
                  },
                  "cold_end_to_end_first_body_byte_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.4355404380185064,
                    "p95": 0.5690667329850838
                  },
                  "cold_end_to_end_response_headers_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.4354242920235265,
                    "p95": 0.5689299000136088
                  },
                  "cold_end_to_end_semantic_ttft_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.793651020998368,
                    "p95": 0.9779021225287579
                  },
                  "cold_end_to_end_stream_total_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.9667132294853218,
                    "p95": 1.8331581669684933
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 37.0,
                    "p95": 39.55
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 1.395e-05,
                    "p95": 1.4399999999999998e-05
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 14.0,
                    "p95": 14.0
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.9297308755049016,
                    "p95": 1.8025811024388503
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.40371929202228785,
                    "p95": 0.5377796893997584
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.4036018130136654,
                    "p95": 0.5376428564282832
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.7626268334861379,
                    "p95": 0.9463682267582043
                  },
                  "setup_dns_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.0012215210008434951,
                    "p95": 0.0016650413163006301
                  },
                  "setup_tcp_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.01148339599603787,
                    "p95": 0.014478962498833422
                  },
                  "setup_tls_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.017997916496824473,
                    "p95": 0.02259496236511039
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 84.86000426935587,
                    "p95": 122.97578848982381
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 51.0,
                    "p95": 54.0
                  }
                }
              },
              "warm": {
                "availability": {
                  "attempted": 30,
                  "rate": 1.0,
                  "successes": 30,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.886482908609522
                  }
                },
                "denominators": {
                  "attempted": 30,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 30,
                  "scheduled": 30,
                  "success": 30
                },
                "metrics": {
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "charged_cost_usd": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 2.835e-05,
                    "p95": 3.0300000000000005e-05
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 38.0,
                    "p95": 40.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 1.41e-05,
                    "p95": 1.5457499999999996e-05
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 14.0,
                    "p95": 16.19999999999999
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.8962873334821779,
                    "p95": 1.0960268963506674
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.36147922949749045,
                    "p95": 0.43618071058008334
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.3613480414787773,
                    "p95": 0.43604239198320993
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.7036326875095256,
                    "p95": 0.9330836377979717
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 85.26395326652326,
                    "p95": 131.6394817915579
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 52.0,
                    "p95": 54.55
                  }
                }
              }
            },
            "role": "gateway"
          },
          {
            "arm_id": "openrouter-openai",
            "baseline": false,
            "conditions": {
              "cold": {
                "availability": {
                  "attempted": 30,
                  "rate": 1.0,
                  "successes": 30,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.886482908609522
                  }
                },
                "denominators": {
                  "attempted": 30,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 30,
                  "scheduled": 30,
                  "success": 30
                },
                "metrics": {
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 0,
                      "ratio": 0.0,
                      "total": 30
                    },
                    "p50": null,
                    "p95": null
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "charged_cost_usd": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 1.575e-05,
                    "p95": 1.61325e-05
                  },
                  "cold_end_to_end_first_body_byte_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.499478020996321,
                    "p95": 0.6956120435992488
                  },
                  "cold_end_to_end_response_headers_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.48873749951599166,
                    "p95": 0.6954786559566855
                  },
                  "cold_end_to_end_semantic_ttft_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.5005642494943459,
                    "p95": 0.7029175142524763
                  },
                  "cold_end_to_end_stream_total_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.6901099374808837,
                    "p95": 0.8744508600619156
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 37.0,
                    "p95": 39.55
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 1.575e-05,
                    "p95": 1.61325e-05
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 17.0,
                    "p95": 17.0
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.6617980415176135,
                    "p95": 0.8448033626715187
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.46975895849755034,
                    "p95": 0.667759970485349
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.459018437017221,
                    "p95": 0.6676265828427856
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.4708451869955752,
                    "p95": 0.6750654411385766
                  },
                  "setup_dns_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.0011508960160426795,
                    "p95": 0.0015771440521348263
                  },
                  "setup_tcp_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.0111028749961406,
                    "p95": 0.011712333545438013
                  },
                  "setup_tls_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.015560562023892999,
                    "p95": 0.020372152770869435
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 100.57065537140309,
                    "p95": 156.85007680269499
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 54.0,
                    "p95": 56.55
                  }
                }
              },
              "warm": {
                "availability": {
                  "attempted": 30,
                  "rate": 1.0,
                  "successes": 30,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.886482908609522
                  }
                },
                "denominators": {
                  "attempted": 30,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 30,
                  "scheduled": 30,
                  "success": 30
                },
                "metrics": {
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 0,
                      "ratio": 0.0,
                      "total": 30
                    },
                    "p50": null,
                    "p95": null
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "charged_cost_usd": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 3.135e-05,
                    "p95": 3.24e-05
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 38.0,
                    "p95": 40.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 1.56e-05,
                    "p95": 1.62e-05
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 17.0,
                    "p95": 17.0
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.5871571250027046,
                    "p95": 0.7637681001157034
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.43211764548323117,
                    "p95": 0.6315843268821475
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.4233607085188851,
                    "p95": 0.4849301542155444
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.4328554795065429,
                    "p95": 0.6547882412996835
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 105.94427864870052,
                    "p95": 415.5487273251547
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 54.5,
                    "p95": 57.0
                  }
                }
              }
            },
            "role": "gateway"
          },
          {
            "arm_id": "vercel-openai",
            "baseline": false,
            "conditions": {
              "cold": {
                "availability": {
                  "attempted": 30,
                  "rate": 1.0,
                  "successes": 30,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.886482908609522
                  }
                },
                "denominators": {
                  "attempted": 30,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 30,
                  "scheduled": 30,
                  "success": 30
                },
                "metrics": {
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "charged_cost_usd": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 1.395e-05,
                    "p95": 1.4399999999999998e-05
                  },
                  "cold_end_to_end_first_body_byte_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.8351534999965224,
                    "p95": 2.5231243521120583
                  },
                  "cold_end_to_end_response_headers_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.8292382500076201,
                    "p95": 2.5188931835087667
                  },
                  "cold_end_to_end_semantic_ttft_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.8415064374858048,
                    "p95": 2.5277202040539075
                  },
                  "cold_end_to_end_stream_total_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 1.004722916521132,
                    "p95": 2.752883272795584
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 37.0,
                    "p95": 39.55
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 1.395e-05,
                    "p95": 1.4399999999999998e-05
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 14.0,
                    "p95": 14.0
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.9506968329951633,
                    "p95": 2.719361543457485
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.7828884375048801,
                    "p95": 2.489602622773959
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.7733296249934938,
                    "p95": 2.4853714541706675
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.7888890624744818,
                    "p95": 2.4941984747158084
                  },
                  "setup_dns_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.001817666518036276,
                    "p95": 0.14762862310744762
                  },
                  "setup_tcp_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.005584062513662502,
                    "p95": 0.006914308253908529
                  },
                  "setup_tls_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.024053103988990188,
                    "p95": 0.028209633677033705
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 87.90931630091568,
                    "p95": 345.2667300252649
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 51.0,
                    "p95": 54.0
                  }
                }
              },
              "warm": {
                "availability": {
                  "attempted": 30,
                  "rate": 1.0,
                  "successes": 30,
                  "wilson95": {
                    "confidence": 0.95,
                    "high": 1.0,
                    "low": 0.886482908609522
                  }
                },
                "denominators": {
                  "attempted": 30,
                  "request_failed": 0,
                  "route_failed": 0,
                  "route_unverifiable": 0,
                  "route_verified": 30,
                  "scheduled": 30,
                  "success": 30
                },
                "metrics": {
                  "cache_write_input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "cached_input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.0,
                    "p95": 0.0
                  },
                  "charged_cost_usd": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 2.835e-05,
                    "p95": 3.0382500000000003e-05
                  },
                  "input_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 38.0,
                    "p95": 40.0
                  },
                  "measured_cost_usd": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 1.4175e-05,
                    "p95": 1.62825e-05
                  },
                  "output_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 14.0,
                    "p95": 18.0
                  },
                  "request_stream_total_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.7264135834993795,
                    "p95": 2.4583423890871896
                  },
                  "request_to_first_body_byte_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.5721195829974022,
                    "p95": 1.6005666395387363
                  },
                  "request_to_response_headers_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.5588160619954579,
                    "p95": 1.5710355376475476
                  },
                  "request_to_semantic_ttft_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 0.5721195829974022,
                    "p95": 1.6050971270393333
                  },
                  "throughput_tokens_per_s": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 94.37047761218236,
                    "p95": 449.19378363279077
                  },
                  "total_tokens": {
                    "coverage": {
                      "covered": 30,
                      "ratio": 1.0,
                      "total": 30
                    },
                    "p50": 52.5,
                    "p95": 55.0
                  }
                }
              }
            },
            "role": "gateway"
          }
        ],
        "baseline_arm_id": "direct-openai",
        "complete_blocks": {
          "cold": 30,
          "warm": 30
        },
        "completion_integrity": null,
        "contrasts": [
          {
            "arm_id": "cloudflare-openai",
            "conditions": {
              "cold": {
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "charged_cost_usd": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cold_end_to_end_first_body_byte_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.12170043648802675,
                    "low": 0.06726702000014484,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.09042116702767089
                },
                "cold_end_to_end_response_headers_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.1216524159826804,
                    "low": 0.06718133254471469,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.09037456254009157
                },
                "cold_end_to_end_semantic_ttft_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.1595789585262537,
                    "low": 0.007124416501028463,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.09380425000563264
                },
                "cold_end_to_end_stream_total_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.18731377049698494,
                    "low": 0.011488624528283253,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.11504197953036055
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.1890914993367912,
                    "low": 0.015864896005950868,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.11766906300908886
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.11944389549898915,
                    "low": 0.07481614596326835,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0899678754794877
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.12039591697975993,
                    "low": 0.07240795850520954,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.08992870850488544
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.1626327504927758,
                    "low": 0.011061979486839846,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.09514824996585958
                },
                "setup_dns_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -0.0001253119553439319,
                    "low": -0.0003124999930150807,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.00027477103867568076
                },
                "setup_tcp_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00021356251090765,
                    "low": -0.0003019794530700892,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -1.8708029529079795e-05
                },
                "setup_tls_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0019422090263105929,
                    "low": -0.0012832079664804041,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.00019310397328808904
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 20.860273527171852,
                    "low": -20.957597029407538,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 1.9003459749691913
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                }
              },
              "warm": {
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "charged_cost_usd": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.12259283295134082,
                    "low": -0.012794332986231893,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.05114497951581143
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.11200035398360342,
                    "low": 0.051706875558011234,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.07910862501012161
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.11193162450217642,
                    "low": 0.04981233406579122,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.08056397899053991
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.11582944041074303,
                    "low": 0.0025938330218195915,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.037403958005597815
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 24.219404585049492,
                    "low": -20.987276097366294,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.2077389019195195
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                }
              }
            },
            "direct_arm": "direct-openai"
          },
          {
            "arm_id": "concentrate-openai",
            "conditions": {
              "cold": {
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "charged_cost_usd": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cold_end_to_end_first_body_byte_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.19392779149347916,
                    "low": 0.14470543750212528,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.16657393801142462
                },
                "cold_end_to_end_response_headers_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.18906864503514953,
                    "low": 0.1446818329859525,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.1665636670368258
                },
                "cold_end_to_end_semantic_ttft_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.23637485451763496,
                    "low": 0.08468770905165002,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.19116072947508655
                },
                "cold_end_to_end_stream_total_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.2920525830122642,
                    "low": 0.05204991699429229,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.19803156197303906
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.27374437503749505,
                    "low": 0.06171727098990232,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.19892531199730001
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.1927400419372134,
                    "low": 0.14380220850580372,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.16858725002384745
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.1900320767555964,
                    "low": 0.1426413330482319,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.16857697904924862
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.23524766700575128,
                    "low": 0.0965253540489357,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.19317883349140175
                },
                "setup_dns_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -0.00024479100829921663,
                    "low": -0.000430583517299965,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.00032902153907343745
                },
                "setup_tcp_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0005222299951128662,
                    "low": -0.00036754150642082095,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.000377853459212929
                },
                "setup_tls_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0027130624803248793,
                    "low": -0.00022147901472635567,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0011074999929405749
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 27.941385781837674,
                    "low": -32.401504564727,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 3.817695001056485
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                }
              },
              "warm": {
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "charged_cost_usd": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.2706499999621883,
                    "low": 0.156668666517362,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.20070966650382616
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.20377760447445326,
                    "low": 0.14063983474479755,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.16801072948146611
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.20367116702254862,
                    "low": 0.14306247950298712,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.16795933348475955
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.23001316701993346,
                    "low": 0.1422742500435561,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.20233466601348482
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 7.422840232342814,
                    "low": -15.673043016897921,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -5.29159021579887
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                }
              }
            },
            "direct_arm": "direct-openai"
          },
          {
            "arm_id": "openrouter-openai",
            "conditions": {
              "cold": {
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 0,
                    "ratio": 0.0,
                    "total": 30
                  },
                  "interval": null,
                  "median_gateway_minus_direct": null
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "charged_cost_usd": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.7999999999999997e-06,
                    "low": 1.7999999999999997e-06,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 1.7999999999999997e-06
                },
                "cold_end_to_end_first_body_byte_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.2527823124837596,
                    "low": 0.2226975834928453,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.2416072295163758
                },
                "cold_end_to_end_response_headers_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.24588058301014826,
                    "low": 0.20653462497284636,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.23038235399872065
                },
                "cold_end_to_end_semantic_ttft_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -0.009755562525242567,
                    "low": -0.17344358400441706,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.08814597950549796
                },
                "cold_end_to_end_stream_total_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.002021333493757993,
                    "low": -0.2061335839680396,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.09709120853221975
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.7999999999999997e-06,
                    "low": 1.7999999999999997e-06,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 1.7999999999999997e-06
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 3.0,
                    "low": 3.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 3.0
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.014045981059461138,
                    "low": -0.17590204198495485,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.0945279790321365
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.2520663960312959,
                    "low": 0.2309046603484603,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.24342847950174473
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.24409531254786998,
                    "low": 0.2121493960148655,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.23311016647494398
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -0.00858166697435081,
                    "low": -0.17558802096755244,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.08753445904585533
                },
                "setup_dns_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -0.0002869585296139121,
                    "low": -0.0005898955278098583,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.00044333349796943367
                },
                "setup_tcp_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -8.949998300522566e-05,
                    "low": -0.0006410834612324834,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.0003782699932344258
                },
                "setup_tls_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00126479149912484,
                    "low": -0.0020909784943796694,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.0005678745219483972
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 47.94634096848077,
                    "low": -2.0912354822923547,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 27.41498365396285
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 3.0,
                    "low": 3.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 3.0
                }
              },
              "warm": {
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 0,
                    "ratio": 0.0,
                    "total": 30
                  },
                  "interval": null,
                  "median_gateway_minus_direct": null
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "charged_cost_usd": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 3.5999999999999994e-06,
                    "low": 2.6999999999999996e-06,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 3.0000000000000018e-06
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 1.7999999999999997e-06,
                    "low": 1.2000000000000004e-06,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 1.7999999999999997e-06
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 3.0,
                    "low": 2.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 3.0
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.005193853547098115,
                    "low": -0.12351249950006604,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.055522812996059656
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.2694512909511104,
                    "low": 0.2287311260588467,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.2501706040056888
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.25157781250891276,
                    "low": 0.2024219169979915,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.2296486044942867
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -0.018306938029127195,
                    "low": -0.1089712499815505,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.054442000488052145
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 27.1157458830417,
                    "low": 4.206847860683574,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 16.020852197066944
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 3.0,
                    "low": 2.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 3.0
                }
              }
            },
            "direct_arm": "direct-openai"
          },
          {
            "arm_id": "vercel-openai",
            "conditions": {
              "cold": {
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "charged_cost_usd": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cold_end_to_end_first_body_byte_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.6373836660059169,
                    "low": 0.4620176244643517,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.5194777495053131
                },
                "cold_end_to_end_response_headers_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.6328783954959363,
                    "low": 0.4505084067335702,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.5110465420002583
                },
                "cold_end_to_end_semantic_ttft_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.29971533350180835,
                    "low": 0.17978658198262565,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.2288713965099305
                },
                "cold_end_to_end_stream_total_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.3379475625115447,
                    "low": 0.16034895749180578,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.238228269998217
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.29213631301536225,
                    "low": 0.14674216747516766,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.23466897898470052
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.6006696879921947,
                    "low": 0.45213562549906783,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.5163948324916419
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.5961508960172068,
                    "low": 0.43974770803470165,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.5079636249865871
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.2967062824573074,
                    "low": 0.17082722901250236,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.22735554198152386
                },
                "setup_dns_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.00042562547605484724,
                    "low": -5.208348738960922e-05,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.00022552150767296553
                },
                "setup_tcp_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": -0.005180478998227045,
                    "low": -0.006166687497170642,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": -0.005376624991185963
                },
                "setup_tls_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.009415479988092557,
                    "low": 0.005901666474528611,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.008299104461912066
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 32.281512534957905,
                    "low": -8.347501492040571,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 9.449165425328447
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                }
              },
              "warm": {
                "cache_write_input_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "cached_input_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "charged_cost_usd": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "input_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "measured_cost_usd": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "output_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                },
                "request_stream_total_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.22041020850883797,
                    "low": -0.04030391649575904,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.07145018698065542
                },
                "request_to_first_body_byte_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.4653479169937782,
                    "low": 0.33173170901136473,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.36314587452216074
                },
                "request_to_response_headers_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.46518966602161527,
                    "low": 0.32805150002241135,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.3480733534961473
                },
                "request_to_semantic_ttft_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.21406552052940242,
                    "low": 0.013774916500551626,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.05843802049639635
                },
                "throughput_tokens_per_s": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 31.988273684160433,
                    "low": -12.96412002218355,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 10.75924396155878
                },
                "total_tokens": {
                  "coverage": {
                    "covered": 30,
                    "ratio": 1.0,
                    "total": 30
                  },
                  "interval": {
                    "confidence": 0.95,
                    "high": 0.0,
                    "low": 0.0,
                    "replicates": 2000
                  },
                  "median_gateway_minus_direct": 0.0
                }
              }
            },
            "direct_arm": "direct-openai"
          }
        ],
        "date": "2026-07-27",
        "default": false,
        "experiment_digest": "e3649dbc6dc423c5209cdc09c73f0d1275bfa6c231a5fa8300e444dc89e6f943",
        "experiment_id": "gpt-4o-mini-five-way-managed-probe-r30-2026-07-27",
        "experiment_path": "gateway-probe/2026-07-27-gpt4o-mini-managed-30/experiment.json",
        "family": "gateway_probe",
        "id": "2026-07-27-gpt4o-mini-managed-30",
        "kind": "release",
        "link": null,
        "model": "GPT-4o mini",
        "model_match": "rolling_alias",
        "output_token_limit_equalities": null,
        "path": null,
        "price_digest": "56ef4049554cc2dc3cb79debc2f5c49d503cfa97c5744d35ffe912815179aea5",
        "result_count": 300,
        "result_schema_version": 3,
        "results_path": "gateway-probe/2026-07-27-gpt4o-mini-managed-30/results.jsonl",
        "retry_summary": null,
        "schedule_digest": "e6b8e2ac79abf1e3b6c6b4ce6b691747a1f9f6bf76c3c78bdb00051e99c097be",
        "scheduled_blocks_per_condition": 30,
        "submitter": "OpenBench",
        "title": "GPT-4o mini Gateway Bench",
        "track": "request_probe",
        "verified_with_commit": "6d1de84d6c96430c47b50a51bd802a986db5c2ba"
      }
    ],
    "note": "Gateway Bench measures request-level transport and serving telemetry. Cold and warm denominators are separate and are never merged.",
    "skipped": [
      {
        "id": "2026-07-28-kimi-k3-managed-30",
        "kind": "gateway_probe",
        "reason": "bundle verification failed: public manifest does not match schema"
      }
    ]
  },
  "generated_by": "obench site",
  "harness": {
    "bundle_count": 5,
    "bundles": [
      {
        "also_seen_as": [],
        "arms": [
          {
            "cost_per_solve_usd": null,
            "fresh_tokens_per_solve": 5518.583333333333,
            "harness": "cursor",
            "median_wall_s": 42.908500000000004,
            "model": "gpt-5.6-sol",
            "n": 42,
            "solve_rate": 0.8571428571428571,
            "solved": 36,
            "token_telemetry_bases": [
              "harness_reported"
            ],
            "token_telemetry_coverage": {
              "covered_rows": 42,
              "native_covered_rows": 42,
              "proxy_covered_rows": 0,
              "ratio": 1.0,
              "total_rows": 42
            },
            "token_telemetry_source": "native",
            "tokens_cache_read_per_solve": 407067.22222222225,
            "tokens_cache_write_per_solve": 45776.333333333336,
            "tokens_input_uncached_per_solve": 72.13888888888889,
            "tokens_output_per_solve": 5446.444444444444,
            "usage_evidence_labels": [],
            "usage_ranking_exclusions": [],
            "wilson95": [
              0.7215861764782122,
              0.9328412431581007
            ]
          },
          {
            "cost_per_solve_usd": null,
            "fresh_tokens_per_solve": null,
            "harness": "devin",
            "median_wall_s": 68.4285,
            "model": "gpt-5.6-sol",
            "n": 42,
            "solve_rate": 0.8095238095238095,
            "solved": 34,
            "token_telemetry_bases": [],
            "token_telemetry_coverage": {
              "covered_rows": 0,
              "native_covered_rows": 0,
              "proxy_covered_rows": 0,
              "ratio": 0.0,
              "total_rows": 42
            },
            "token_telemetry_source": null,
            "tokens_cache_read_per_solve": null,
            "tokens_cache_write_per_solve": null,
            "tokens_input_uncached_per_solve": null,
            "tokens_output_per_solve": null,
            "usage_evidence_labels": [],
            "usage_ranking_exclusions": [],
            "wilson95": [
              0.6669893264500356,
              0.9001811039014356
            ]
          },
          {
            "cost_per_solve_usd": null,
            "fresh_tokens_per_solve": 46215.58823529412,
            "harness": "grokbuild",
            "median_wall_s": 86.2995,
            "model": "gpt-5.6-sol",
            "n": 42,
            "solve_rate": 0.8095238095238095,
            "solved": 34,
            "token_telemetry_bases": [
              "proxy_measured"
            ],
            "token_telemetry_coverage": {
              "covered_rows": 42,
              "native_covered_rows": 42,
              "proxy_covered_rows": 42,
              "ratio": 1.0,
              "total_rows": 42
            },
            "token_telemetry_source": "proxy",
            "tokens_cache_read_per_solve": 368079.0588235294,
            "tokens_cache_write_per_solve": 0.0,
            "tokens_input_uncached_per_solve": 38726.82352941176,
            "tokens_output_per_solve": 7488.764705882353,
            "usage_evidence_labels": [],
            "usage_ranking_exclusions": [],
            "wilson95": [
              0.6669893264500356,
              0.9001811039014356
            ]
          },
          {
            "cost_per_solve_usd": null,
            "fresh_tokens_per_solve": 42634.5,
            "harness": "opencode",
            "median_wall_s": 60.524,
            "model": "gpt-5.6-sol",
            "n": 42,
            "solve_rate": 0.8095238095238095,
            "solved": 34,
            "token_telemetry_bases": [
              "vendor_split"
            ],
            "token_telemetry_coverage": {
              "covered_rows": 42,
              "native_covered_rows": 42,
              "proxy_covered_rows": 0,
              "ratio": 1.0,
              "total_rows": 42
            },
            "token_telemetry_source": "native",
            "tokens_cache_read_per_solve": 313916.23529411765,
            "tokens_cache_write_per_solve": 0.0,
            "tokens_input_uncached_per_solve": 36707.55882352941,
            "tokens_output_per_solve": 5926.941176470588,
            "usage_evidence_labels": [],
            "usage_ranking_exclusions": [],
            "wilson95": [
              0.6669893264500356,
              0.9001811039014356
            ]
          },
          {
            "cost_per_solve_usd": null,
            "fresh_tokens_per_solve": 40255.0,
            "harness": "claude",
            "median_wall_s": 58.793,
            "model": "gpt-5.6-sol",
            "n": 42,
            "solve_rate": 0.7619047619047619,
            "solved": 32,
            "token_telemetry_bases": [
              "proxy_measured"
            ],
            "token_telemetry_coverage": {
              "covered_rows": 42,
              "native_covered_rows": 42,
              "proxy_covered_rows": 42,
              "ratio": 1.0,
              "total_rows": 42
            },
            "token_telemetry_source": "proxy",
            "tokens_cache_read_per_solve": 153488.0,
            "tokens_cache_write_per_solve": 0.0,
            "tokens_input_uncached_per_solve": 34455.125,
            "tokens_output_per_solve": 5799.875,
            "usage_evidence_labels": [],
            "usage_ranking_exclusions": [],
            "wilson95": [
              0.6147216491768112,
              0.8651917918898182
            ]
          },
          {
            "cost_per_solve_usd": null,
            "fresh_tokens_per_solve": 37602.0625,
            "harness": "pi",
            "median_wall_s": 39.670500000000004,
            "model": "gpt-5.6-sol",
            "n": 42,
            "solve_rate": 0.7619047619047619,
            "solved": 32,
            "token_telemetry_bases": [
              "proxy_measured"
            ],
            "token_telemetry_coverage": {
              "covered_rows": 42,
              "native_covered_rows": 42,
              "proxy_covered_rows": 42,
              "ratio": 1.0,
              "total_rows": 42
            },
            "token_telemetry_source": "proxy",
            "tokens_cache_read_per_solve": 111408.0,
            "tokens_cache_write_per_solve": 0.0,
            "tokens_input_uncached_per_solve": 32915.71875,
            "tokens_output_per_solve": 4686.34375,
            "usage_evidence_labels": [],
            "usage_ranking_exclusions": [],
            "wilson95": [
              0.6147216491768112,
              0.8651917918898182
            ]
          },
          {
            "cost_per_solve_usd": null,
            "fresh_tokens_per_solve": 117107.35483870968,
            "harness": "codex",
            "median_wall_s": 94.603,
            "model": "gpt-5.6-sol",
            "n": 42,
            "solve_rate": 0.7380952380952381,
            "solved": 31,
            "token_telemetry_bases": [
              "proxy_measured"
            ],
            "token_telemetry_coverage": {
              "covered_rows": 42,
              "native_covered_rows": 39,
              "proxy_covered_rows": 42,
              "ratio": 1.0,
              "total_rows": 42
            },
            "token_telemetry_source": "proxy",
            "tokens_cache_read_per_solve": 1164510.9677419355,
            "tokens_cache_write_per_solve": 0.0,
            "tokens_input_uncached_per_solve": 102174.32258064517,
            "tokens_output_per_solve": 14933.032258064517,
            "usage_evidence_labels": [],
            "usage_ranking_exclusions": [],
            "wilson95": [
              0.5893098001502446,
              0.8469751462739642
            ]
          }
        ],
        "caveats": [
          "Archived task definitions differ from or are unavailable in the current checkout for add-feature, make-ci-green, taskflow, terminal-bench/schemelike-metacircular-eval, webcore. The result seal and bundled evidence still verify; recorded task digests remain the run-time fingerprints."
        ],
        "claim": null,
        "countable_rows": 312,
        "date": "2026-07-21",
        "excluded_counts": {
          "infra": 2,
          "rate_limited": 1
        },
        "family": "harness",
        "has_caveats": true,
        "id": "2026-07-21-gpt56",
        "kind": "release",
        "link": null,
        "matched": {
          "comparison_identity": "legacy_task_trial",
          "duplicate_cells_excluded": 0,
          "duplicate_rows_excluded": 0,
          "groups_compared": 7,
          "matched_cells_per_group": 42,
          "matched_rows": 294,
          "unmatched_countable_rows": 18
        },
        "models": [
          "gpt-5.6-sol"
        ],
        "path": "releases/2026-07-21-gpt56/index.html",
        "raw_rows": 315,
        "results_path": "releases/2026-07-21-gpt56/results.jsonl",
        "results_sha256": "546f167576aac07d8d4c889d559ac86d9f9ad427c5fa53e5a421059f9db498b6",
        "submitter": null,
        "table": "matched",
        "task_set_digest": "d949d46e8aae81489a91d1925ec14f91e4667f5c69ccb314e8fdc611d966c059",
        "title": "OpenBench: gpt-5.6 across 7 harnesses"
      },
      {
        "also_seen_as": [],
        "arms": [
          {
            "cost_per_solve_usd": null,
            "fresh_tokens_per_solve": 78507.96666666666,
            "harness": "claude",
            "median_wall_s": 47.6005,
            "model": "deepseek-v4-flash",
            "n": 37,
            "solve_rate": 0.8108108108108109,
            "solved": 30,
            "token_telemetry_bases": [
              "proxy_measured"
            ],
            "token_telemetry_coverage": {
              "covered_rows": 37,
              "native_covered_rows": 36,
              "proxy_covered_rows": 37,
              "ratio": 1.0,
              "total_rows": 37
            },
            "token_telemetry_source": "proxy",
            "tokens_cache_read_per_solve": 1918016.0,
            "tokens_cache_write_per_solve": 0.0,
            "tokens_input_uncached_per_solve": 33949.46666666667,
            "tokens_output_per_solve": 44558.5,
            "usage_evidence_labels": [],
            "usage_ranking_exclusions": [],
            "wilson95": [
              0.6579498842904441,
              0.9052014124266288
            ]
          },
          {
            "cost_per_solve_usd": null,
            "fresh_tokens_per_solve": 62945.78571428572,
            "harness": "pi",
            "median_wall_s": 24.4575,
            "model": "deepseek-v4-flash",
            "n": 37,
            "solve_rate": 0.7567567567567568,
            "solved": 28,
            "token_telemetry_bases": [
              "proxy_measured"
            ],
            "token_telemetry_coverage": {
              "covered_rows": 37,
              "native_covered_rows": 34,
              "proxy_covered_rows": 37,
              "ratio": 1.0,
              "total_rows": 37
            },
            "token_telemetry_source": "proxy",
            "tokens_cache_read_per_solve": 1601773.7142857143,
            "tokens_cache_write_per_solve": 0.0,
            "tokens_input_uncached_per_solve": 33190.0,
            "tokens_output_per_solve": 29755.785714285714,
            "usage_evidence_labels": [],
            "usage_ranking_exclusions": [],
            "wilson95": [
              0.5988239750929106,
              0.8663879656733671
            ]
          },
          {
            "cost_per_solve_usd": null,
            "fresh_tokens_per_solve": 68353.57692307692,
            "harness": "opencode",
            "median_wall_s": 40.271,
            "model": "deepseek-v4-flash",
            "n": 37,
            "solve_rate": 0.7027027027027027,
            "solved": 26,
            "token_telemetry_bases": [
              "proxy_measured"
            ],
            "token_telemetry_coverage": {
              "covered_rows": 37,
              "native_covered_rows": 36,
              "proxy_covered_rows": 37,
              "ratio": 1.0,
              "total_rows": 37
            },
            "token_telemetry_source": "proxy",
            "tokens_cache_read_per_solve": 1529383.3846153845,
            "tokens_cache_write_per_solve": 0.0,
            "tokens_input_uncached_per_solve": 37065.346153846156,
            "tokens_output_per_solve": 31288.23076923077,
            "usage_evidence_labels": [],
            "usage_ranking_exclusions": [],
            "wilson95": [
              0.5421656082795642,
              0.8251069765359181
            ]
          },
          {
            "cost_per_solve_usd": null,
            "fresh_tokens_per_solve": 120124.44,
            "harness": "grokbuild",
            "median_wall_s": 27.67,
            "model": "deepseek-v4-flash",
            "n": 37,
            "solve_rate": 0.6756756756756757,
            "solved": 25,
            "token_telemetry_bases": [
              "proxy_measured"
            ],
            "token_telemetry_coverage": {
              "covered_rows": 37,
              "native_covered_rows": 37,
              "proxy_covered_rows": 37,
              "ratio": 1.0,
              "total_rows": 37
            },
            "token_telemetry_source": "proxy",
            "tokens_cache_read_per_solve": 1579745.28,
            "tokens_cache_write_per_solve": 0.0,
            "tokens_input_uncached_per_solve": 74114.76,
            "tokens_output_per_solve": 46009.68,
            "usage_evidence_labels": [],
            "usage_ranking_exclusions": [],
            "wilson95": [
              0.5146336850482394,
              0.8036692217918452
            ]
          },
          {
            "cost_per_solve_usd": null,
            "fresh_tokens_per_solve": 86414.16666666667,
            "harness": "codex",
            "median_wall_s": 27.182,
            "model": "deepseek-v4-flash",
            "n": 37,
            "solve_rate": 0.6486486486486487,
            "solved": 24,
            "token_telemetry_bases": [
              "proxy_measured"
            ],
            "token_telemetry_coverage": {
              "covered_rows": 37,
              "native_covered_rows": 32,
              "proxy_covered_rows": 37,
              "ratio": 1.0,
              "total_rows": 37
            },
            "token_telemetry_source": "proxy",
            "tokens_cache_read_per_solve": 1119568.0,
            "tokens_cache_write_per_solve": 0.0,
            "tokens_input_uncached_per_solve": 36972.041666666664,
            "tokens_output_per_solve": 49442.125,
            "usage_evidence_labels": [],
            "usage_ranking_exclusions": [],
            "wilson95": [
              0.48758729517569566,
              0.7817459336889914
            ]
          }
        ],
        "caveats": [
          "Archived first-party matrix published without rerunning models. The source contains 225 rows across 5 harnesses, 15 tasks, and 3 trials.",
          "The bundled results SHA is sealed. Current-tree task checks report expected post-run definition drift; recorded task digests are retained as run-time fingerprints.",
          "The unified board ranks only task/trial cells shared by every harness. The historical release page retains its older per-arm countable view, so its denominators differ.",
          "Harness and host/container version stamps vary in the archive. Token columns use the complete counting-proxy lane for every displayed harness.",
          "Archived task definitions differ from or are unavailable in the current checkout for add-feature, make-ci-green, taskflow, terminal-bench/feal-differential-cryptanalysis, terminal-bench/llm-inference-batching-scheduler, terminal-bench/schemelike-metacircular-eval, webcore. The result seal and bundled evidence still verify; recorded task digests remain the run-time fingerprints."
        ],
        "claim": null,
        "countable_rows": 215,
        "date": "2026-07-20",
        "excluded_counts": {
          "infra": 8,
          "rate_limited": 2
        },
        "family": "harness",
        "has_caveats": true,
        "id": "2026-07-20-deepseek",
        "kind": "release",
        "link": null,
        "matched": {
          "comparison_identity": "legacy_task_trial",
          "duplicate_cells_excluded": 0,
          "duplicate_rows_excluded": 0,
          "groups_compared": 5,
          "matched_cells_per_group": 37,
          "matched_rows": 185,
          "unmatched_countable_rows": 30
        },
        "models": [
          "deepseek-v4-flash"
        ],
        "path": "releases/2026-07-20-deepseek/index.html",
        "raw_rows": 225,
        "results_path": "releases/2026-07-20-deepseek/results.jsonl",
        "results_sha256": "ec6b9d58fd3aa6d001db67dc8333e1590c7b1acaa65595f05347413ad23ef414",
        "submitter": null,
        "table": "matched",
        "task_set_digest": "1c4805e7fcd0fd8ff18912c3f745ad243edd2dfd2448f933efe5d0a79966f5bc",
        "title": "OpenBench: deepseek across 5 harnesses"
      },
      {
        "also_seen_as": [],
        "arms": [
          {
            "cost_per_solve_usd": null,
            "fresh_tokens_per_solve": 36706.57142857143,
            "harness": "cursor",
            "median_wall_s": 25.938,
            "model": "grok-4.5",
            "n": 39,
            "solve_rate": 0.8974358974358975,
            "solved": 35,
            "token_telemetry_bases": [
              "harness_reported"
            ],
            "token_telemetry_coverage": {
              "covered_rows": 39,
              "native_covered_rows": 39,
              "proxy_covered_rows": 0,
              "ratio": 1.0,
              "total_rows": 39
            },
            "token_telemetry_source": "native",
            "tokens_cache_read_per_solve": 247504.45714285714,
            "tokens_cache_write_per_solve": 0.0,
            "tokens_input_uncached_per_solve": 28600.485714285714,
            "tokens_output_per_solve": 8106.085714285714,
            "usage_evidence_labels": [],
            "usage_ranking_exclusions": [],
            "wilson95": [
              0.7642084129775517,
              0.9593873444171301
            ]
          },
          {
            "cost_per_solve_usd": null,
            "fresh_tokens_per_solve": 58602.617647058825,
            "harness": "grokbuild",
            "median_wall_s": 42.2565,
            "model": "grok-4.5",
            "n": 39,
            "solve_rate": 0.8717948717948718,
            "solved": 34,
            "token_telemetry_bases": [
              "unknown"
            ],
            "token_telemetry_coverage": {
              "covered_rows": 39,
              "native_covered_rows": 39,
              "proxy_covered_rows": 0,
              "ratio": 1.0,
              "total_rows": 39
            },
            "token_telemetry_source": "native",
            "tokens_cache_read_per_solve": 212931.76470588235,
            "tokens_cache_write_per_solve": null,
            "tokens_input_uncached_per_solve": 49844.58823529412,
            "tokens_output_per_solve": 8758.029411764706,
            "usage_evidence_labels": [],
            "usage_ranking_exclusions": [],
            "wilson95": [
              0.7329395096236953,
              0.9439726505197168
            ]
          },
          {
            "cost_per_solve_usd": null,
            "fresh_tokens_per_solve": null,
            "harness": "pi",
            "median_wall_s": 25.284,
            "model": "grok-4.5",
            "n": 39,
            "solve_rate": 0.7948717948717948,
            "solved": 31,
            "token_telemetry_bases": [],
            "token_telemetry_coverage": {
              "covered_rows": 0,
              "native_covered_rows": 38,
              "proxy_covered_rows": 0,
              "ratio": 0.0,
              "total_rows": 39
            },
            "token_telemetry_source": null,
            "tokens_cache_read_per_solve": null,
            "tokens_cache_write_per_solve": null,
            "tokens_input_uncached_per_solve": null,
            "tokens_output_per_solve": null,
            "usage_evidence_labels": [],
            "usage_ranking_exclusions": [],
            "wilson95": [
              0.6446572703687137,
              0.8922040980208888
            ]
          },
          {
            "cost_per_solve_usd": null,
            "fresh_tokens_per_solve": 55008.13333333333,
            "harness": "opencode",
            "median_wall_s": 21.2275,
            "model": "grok-4.5",
            "n": 39,
            "solve_rate": 0.7692307692307693,
            "solved": 30,
            "token_telemetry_bases": [
              "vendor_split"
            ],
            "token_telemetry_coverage": {
              "covered_rows": 39,
              "native_covered_rows": 39,
              "proxy_covered_rows": 0,
              "ratio": 1.0,
              "total_rows": 39
            },
            "token_telemetry_source": "native",
            "tokens_cache_read_per_solve": 250734.93333333332,
            "tokens_cache_write_per_solve": 0.0,
            "tokens_input_uncached_per_solve": 45493.666666666664,
            "tokens_output_per_solve": 9514.466666666667,
            "usage_evidence_labels": [],
            "usage_ranking_exclusions": [],
            "wilson95": [
              0.6166343954261705,
              0.8735433757121625
            ]
          }
        ],
        "caveats": [
          "Archived first-party matrix published without rerunning models. The source contains 180 rows across 4 harnesses, 15 tasks, and 3 trials.",
          "The bundled results SHA is sealed. Current-tree task checks report expected post-run definition drift; recorded task digests are retained as run-time fingerprints.",
          "The unified board ranks only task/trial cells shared by every harness. The historical release page retains its older per-arm countable view, so its denominators differ.",
          "No counting-proxy telemetry was retained for this run. Complete native split telemetry is available only where the board reports 100% native coverage.",
          "Archived task definitions differ from or are unavailable in the current checkout for add-feature, make-ci-green, taskflow, terminal-bench/feal-differential-cryptanalysis, terminal-bench/llm-inference-batching-scheduler, terminal-bench/schemelike-metacircular-eval, webcore. The result seal and bundled evidence still verify; recorded task digests remain the run-time fingerprints."
        ],
        "claim": null,
        "countable_rows": 174,
        "date": "2026-07-20",
        "excluded_counts": {
          "infra": 6
        },
        "family": "harness",
        "has_caveats": true,
        "id": "2026-07-20-grok45",
        "kind": "release",
        "link": null,
        "matched": {
          "comparison_identity": "legacy_task_trial",
          "duplicate_cells_excluded": 0,
          "duplicate_rows_excluded": 0,
          "groups_compared": 4,
          "matched_cells_per_group": 39,
          "matched_rows": 156,
          "unmatched_countable_rows": 18
        },
        "models": [
          "grok-4.5"
        ],
        "path": "releases/2026-07-20-grok45/index.html",
        "raw_rows": 180,
        "results_path": "releases/2026-07-20-grok45/results.jsonl",
        "results_sha256": "3905ac339eea53b2e66fd2eb6b3a670174d364a40111777d4428619aba3f0081",
        "submitter": null,
        "table": "matched",
        "task_set_digest": "1c4805e7fcd0fd8ff18912c3f745ad243edd2dfd2448f933efe5d0a79966f5bc",
        "title": "OpenBench: grok-4.5 across 4 harnesses"
      },
      {
        "also_seen_as": [],
        "arms": [
          {
            "cost_per_solve_usd": null,
            "fresh_tokens_per_solve": 5856.0,
            "harness": "aider",
            "median_wall_s": 39.0355,
            "model": "deepseek-v4-flash",
            "n": 8,
            "solve_rate": 1.0,
            "solved": 8,
            "token_telemetry_bases": [
              "proxy_measured"
            ],
            "token_telemetry_coverage": {
              "covered_rows": 8,
              "native_covered_rows": 0,
              "proxy_covered_rows": 8,
              "ratio": 1.0,
              "total_rows": 8
            },
            "token_telemetry_source": "proxy",
            "tokens_cache_read_per_solve": 336.0,
            "tokens_cache_write_per_solve": 0.0,
            "tokens_input_uncached_per_solve": 2307.625,
            "tokens_output_per_solve": 3548.375,
            "usage_evidence_labels": [],
            "usage_ranking_exclusions": [],
            "wilson95": [
              0.6755843804891231,
              1.0
            ]
          },
          {
            "cost_per_solve_usd": null,
            "fresh_tokens_per_solve": 13669.375,
            "harness": "opencode",
            "median_wall_s": 31.720999999999997,
            "model": "deepseek-v4-flash",
            "n": 8,
            "solve_rate": 1.0,
            "solved": 8,
            "token_telemetry_bases": [
              "proxy_measured"
            ],
            "token_telemetry_coverage": {
              "covered_rows": 8,
              "native_covered_rows": 8,
              "proxy_covered_rows": 8,
              "ratio": 1.0,
              "total_rows": 8
            },
            "token_telemetry_source": "proxy",
            "tokens_cache_read_per_solve": 91600.0,
            "tokens_cache_write_per_solve": 0.0,
            "tokens_input_uncached_per_solve": 10327.875,
            "tokens_output_per_solve": 3341.5,
            "usage_evidence_labels": [],
            "usage_ranking_exclusions": [],
            "wilson95": [
              0.6755843804891231,
              1.0
            ]
          },
          {
            "cost_per_solve_usd": null,
            "fresh_tokens_per_solve": 7742.25,
            "harness": "pi",
            "median_wall_s": 26.444,
            "model": "deepseek-v4-flash",
            "n": 8,
            "solve_rate": 1.0,
            "solved": 8,
            "token_telemetry_bases": [
              "proxy_measured"
            ],
            "token_telemetry_coverage": {
              "covered_rows": 8,
              "native_covered_rows": 8,
              "proxy_covered_rows": 8,
              "ratio": 1.0,
              "total_rows": 8
            },
            "token_telemetry_source": "proxy",
            "tokens_cache_read_per_solve": 35456.0,
            "tokens_cache_write_per_solve": 0.0,
            "tokens_input_uncached_per_solve": 4927.75,
            "tokens_output_per_solve": 2814.5,
            "usage_evidence_labels": [],
            "usage_ranking_exclusions": [],
            "wilson95": [
              0.6755843804891231,
              1.0
            ]
          }
        ],
        "caveats": [
          "Harness mode is not apples-to-apples: aider ran as a one-shot --message invocation, while pi and opencode ran full agentic loops. Token comparison reflects that difference in harness mode, not a pure like-for-like agent loop contest.",
          "One aider cell excluded: 1 of 9 aider cells was infrastructure-classified and is excluded from solve-rate denominators (aider reported as 8/8). pi and opencode are 9/9.",
          "Uniform total-token basis: totals include uncached input, output, and cache reads from split fields; the vendor aggregate is never used. Values are 6,192 vs 48,262 vs 111,380 tokens/solve (aider vs pi vs opencode).",
          "Verifiable bundle: this release includes results.jsonl and provenance.json. Re-check digests with obench verify docs/releases/2026-07-20-aider-showcase."
        ],
        "claim": "Seed community submission: the verified aider showcase bundle (same artifact as docs/releases/2026-07-20-aider-showcase).",
        "countable_rows": 26,
        "date": "2026-07-20",
        "excluded_counts": {
          "infra": 1
        },
        "family": "harness",
        "has_caveats": true,
        "id": "openbench-aider-showcase",
        "kind": "community",
        "link": "https://github.com/minghinmatthewlam/openbench/tree/main/docs/releases/2026-07-20-aider-showcase",
        "matched": {
          "comparison_identity": "legacy_task_trial",
          "duplicate_cells_excluded": 0,
          "duplicate_rows_excluded": 0,
          "groups_compared": 3,
          "matched_cells_per_group": 8,
          "matched_rows": 24,
          "unmatched_countable_rows": 2
        },
        "models": [
          "deepseek-v4-flash"
        ],
        "path": "../data/community/openbench-aider-showcase/index.html",
        "raw_rows": 27,
        "results_path": "../data/community/openbench-aider-showcase/results.jsonl",
        "results_sha256": "677642cc0649ecfd42231dbdf49fc5aa0b2c2a8161ccb879358bfeec40e45192",
        "submitter": "openbench",
        "table": "matched",
        "task_set_digest": "8b757f8abf3816cac2ddc80e5ceaefa3f4a323c47c2431258a655d4d3b7f2e43",
        "title": "OpenBench showcase: BYO aider vs pi vs opencode (deepseek-v4-flash)"
      },
      {
        "also_seen_as": [],
        "arms": [
          {
            "cost_per_solve_usd": null,
            "fresh_tokens_per_solve": null,
            "harness": "pi",
            "median_wall_s": 693.4925000000001,
            "model": "glm-5.2",
            "n": 8,
            "solve_rate": 1.0,
            "solved": 8,
            "token_telemetry_bases": [],
            "token_telemetry_coverage": {
              "covered_rows": 0,
              "native_covered_rows": 6,
              "proxy_covered_rows": 0,
              "ratio": 0.0,
              "total_rows": 8
            },
            "token_telemetry_source": null,
            "tokens_cache_read_per_solve": null,
            "tokens_cache_write_per_solve": null,
            "tokens_input_uncached_per_solve": null,
            "tokens_output_per_solve": null,
            "usage_evidence_labels": [],
            "usage_ranking_exclusions": [],
            "wilson95": [
              0.6755843804891231,
              1.0
            ]
          },
          {
            "cost_per_solve_usd": null,
            "fresh_tokens_per_solve": 94197.42857142857,
            "harness": "opencode",
            "median_wall_s": 661.219,
            "model": "glm-5.2",
            "n": 8,
            "solve_rate": 0.875,
            "solved": 7,
            "token_telemetry_bases": [
              "vendor_split"
            ],
            "token_telemetry_coverage": {
              "covered_rows": 8,
              "native_covered_rows": 8,
              "proxy_covered_rows": 0,
              "ratio": 1.0,
              "total_rows": 8
            },
            "token_telemetry_source": "native",
            "tokens_cache_read_per_solve": 973165.7142857143,
            "tokens_cache_write_per_solve": 0.0,
            "tokens_input_uncached_per_solve": 54762.0,
            "tokens_output_per_solve": 39435.42857142857,
            "usage_evidence_labels": [],
            "usage_ranking_exclusions": [],
            "wilson95": [
              0.5291051942301385,
              0.9775830911367039
            ]
          },
          {
            "cost_per_solve_usd": null,
            "fresh_tokens_per_solve": null,
            "harness": "claude",
            "median_wall_s": 568.97,
            "model": "glm-5.2",
            "n": 8,
            "solve_rate": 0.625,
            "solved": 5,
            "token_telemetry_bases": [],
            "token_telemetry_coverage": {
              "covered_rows": 0,
              "native_covered_rows": 4,
              "proxy_covered_rows": 0,
              "ratio": 0.0,
              "total_rows": 8
            },
            "token_telemetry_source": null,
            "tokens_cache_read_per_solve": null,
            "tokens_cache_write_per_solve": null,
            "tokens_input_uncached_per_solve": null,
            "tokens_output_per_solve": null,
            "usage_evidence_labels": [],
            "usage_ranking_exclusions": [],
            "wilson95": [
              0.3057378545838018,
              0.863158240538479
            ]
          },
          {
            "cost_per_solve_usd": null,
            "fresh_tokens_per_solve": null,
            "harness": "grokbuild",
            "median_wall_s": 267.572,
            "model": "glm-5.2",
            "n": 8,
            "solve_rate": 0.625,
            "solved": 5,
            "token_telemetry_bases": [],
            "token_telemetry_coverage": {
              "covered_rows": 0,
              "native_covered_rows": 0,
              "proxy_covered_rows": 0,
              "ratio": 0.0,
              "total_rows": 8
            },
            "token_telemetry_source": null,
            "tokens_cache_read_per_solve": null,
            "tokens_cache_write_per_solve": null,
            "tokens_input_uncached_per_solve": null,
            "tokens_output_per_solve": null,
            "usage_evidence_labels": [],
            "usage_ranking_exclusions": [],
            "wilson95": [
              0.3057378545838018,
              0.863158240538479
            ]
          },
          {
            "cost_per_solve_usd": null,
            "fresh_tokens_per_solve": null,
            "harness": "codex",
            "median_wall_s": 801.077,
            "model": "glm-5.2",
            "n": 8,
            "solve_rate": 0.5,
            "solved": 4,
            "token_telemetry_bases": [],
            "token_telemetry_coverage": {
              "covered_rows": 0,
              "native_covered_rows": 4,
              "proxy_covered_rows": 0,
              "ratio": 0.0,
              "total_rows": 8
            },
            "token_telemetry_source": null,
            "tokens_cache_read_per_solve": null,
            "tokens_cache_write_per_solve": null,
            "tokens_input_uncached_per_solve": null,
            "tokens_output_per_solve": null,
            "usage_evidence_labels": [],
            "usage_ranking_exclusions": [],
            "wilson95": [
              0.2152125268244419,
              0.7847874731755581
            ]
          }
        ],
        "caveats": [
          "Quarantine-safe derived bundle: all 15 terminal-bench/cancel-async-tasks rows were excluded because its load-sensitive checker is binding-quarantined. No model was rerun.",
          "The bundled results SHA is sealed. Current-tree task checks report expected post-run definition drift; recorded task digests are retained as run-time fingerprints.",
          "The retained archive lacks image digests, version-source labels, and an explicit per-row timeout field; contemporaneous version-stamp drift is documented in the bundle README.",
          "The unified board ranks only task/trial cells shared by every harness. The release page also shows the per-arm countable view, so its denominators differ.",
          "Archived task definitions differ from or are unavailable in the current checkout for terminal-bench/feal-differential-cryptanalysis, terminal-bench/llm-inference-batching-scheduler, terminal-bench/schemelike-metacircular-eval. The result seal and bundled evidence still verify; recorded task digests remain the run-time fingerprints."
        ],
        "claim": null,
        "countable_rows": 44,
        "date": "2026-07-09",
        "excluded_counts": {
          "infra": 1
        },
        "family": "harness",
        "has_caveats": true,
        "id": "2026-07-09-glm52",
        "kind": "release",
        "link": null,
        "matched": {
          "comparison_identity": "legacy_task_trial",
          "duplicate_cells_excluded": 0,
          "duplicate_rows_excluded": 0,
          "groups_compared": 5,
          "matched_cells_per_group": 8,
          "matched_rows": 40,
          "unmatched_countable_rows": 4
        },
        "models": [
          "glm-5.2"
        ],
        "path": "releases/2026-07-09-glm52/index.html",
        "raw_rows": 45,
        "results_path": "releases/2026-07-09-glm52/results.jsonl",
        "results_sha256": "5434f3f8aa66a708936cc322e26b23dd5670caa5532163f1c27fe3cb1a57349c",
        "submitter": null,
        "table": "matched",
        "task_set_digest": "46075fa27b81914497dc37b6ff595ef49dd09b04154ac9f9e93c33892aa0168f",
        "title": "OpenBench: GLM 5.2 quarantine-safe archive"
      }
    ],
    "note": "Scores are not comparable across bundles: task sets, trial counts, and timeout caps differ.",
    "skipped": [
      {
        "id": "2026-07-02-m3",
        "kind": "release",
        "path": "releases/2026-07-02-m3/index.html",
        "reason": "no results.jsonl (HTML-only release page)"
      },
      {
        "id": "2026-07-20-kimi-k3",
        "kind": "release",
        "path": "releases/2026-07-20-kimi-k3/index.html",
        "reason": "no results.jsonl (HTML-only release page)"
      }
    ]
  },
  "packs": [
    {
      "content_sha256": "3b1e576398a8c8f6d42a514ce83ec610f81742edd79f733453bbff681dbeca58",
      "description": "Tiny polarity-checked smoke tasks (make-it-run, fix-failing-test)",
      "id": "openbench/core-smoke",
      "kind": "tasks",
      "latest": "1.0.0",
      "license": "Apache-2.0",
      "name": "core-smoke",
      "org": "openbench",
      "source": "data/packs/openbench-core-smoke",
      "tasks": [
        "make-it-run",
        "fix-failing-test"
      ]
    }
  ],
  "releases": [
    {
      "date": "2026-07-02",
      "id": "2026-07-02-m3",
      "models": [
        "gpt-5.5-medium"
      ],
      "path": "releases/2026-07-02-m3/index.html",
      "status": "final",
      "title": "OpenBench M3 — Harness Comparison"
    },
    {
      "date": "2026-07-09",
      "id": "2026-07-09-glm52",
      "models": [
        "glm-5.2"
      ],
      "path": "releases/2026-07-09-glm52/index.html",
      "status": "final",
      "title": "OpenBench: GLM 5.2 quarantine-safe archive"
    },
    {
      "date": "2026-07-20",
      "id": "2026-07-20-grok45",
      "models": [
        "grok-4.5"
      ],
      "path": "releases/2026-07-20-grok45/index.html",
      "status": "final",
      "title": "OpenBench: grok-4.5 across 4 harnesses"
    },
    {
      "date": "2026-07-20",
      "id": "2026-07-20-kimi-k3",
      "models": [
        "kimi-k3"
      ],
      "path": "releases/2026-07-20-kimi-k3/index.html",
      "status": "final",
      "title": "OpenBench: kimi-k3 across 3 harnesses"
    },
    {
      "date": "2026-07-20",
      "id": "2026-07-20-deepseek",
      "models": [
        "deepseek-v4-flash"
      ],
      "path": "releases/2026-07-20-deepseek/index.html",
      "status": "final",
      "title": "OpenBench: deepseek across 5 harnesses"
    },
    {
      "date": "2026-07-21",
      "id": "2026-07-21-gpt56",
      "models": [
        "gpt-5.6-sol"
      ],
      "path": "releases/2026-07-21-gpt56/index.html",
      "status": "final",
      "title": "OpenBench: gpt-5.6 across 7 harnesses"
    }
  ],
  "schema_version": 5,
  "site_metadata": {
    "canonical_url": "https://openbench.run/openbench/",
    "social_image_url": "https://openbench.run/openbench/og.png"
  }
}