{
  "generated_at_utc": "2026-10-02T09:11:51.530602+00:00",
  "analysis_status": "post-hoc exploratory uncertainty supplement; original records, endpoints, labels and scores unchanged",
  "methods": {
    "confidence_level": 0.95,
    "paired_bootstrap_replicates": 20000,
    "interval": "2.5th and 97.5th percentiles, linear interpolation",
    "resampling_unit": "case identity; all compared arms retained together",
    "accuracy_interval": "Wilson score interval with z=1.959963984540054",
    "direction": "Each comparison identifies its left and right arms; differences are left minus right; ratios are left divided by right",
    "independence_limit": "Case-level bootstrap/Wilson calculations do not account for common synthetic templates, duplicated passages, temporal dependence or provider-state variation. They are descriptive intervals for this fixed pilot, not population-wide guarantees.",
    "multiplicity": "No multiplicity adjustment. Main endpoints existed before evaluation; these additional interval calculations and sensitivities are post hoc."
  },
  "integrity": {
    "main": {
      "checks": {
        "status_complete": true,
        "case_hash_matches": true,
        "source_hash_matches": true,
        "phase_counts_match": true,
        "consecutive_sequence_numbers": true,
        "unique_generation_ids": true,
        "http_200_and_valid": true,
        "exact_answer_schema": true,
        "probability_matches_answer": true,
        "logged_answer_matches_raw_response": true,
        "raw_cost_matches_record_and_finite": true,
        "billed_total_matches_metadata": true,
        "positive_finite_latencies": true,
        "targets_and_inputs_match_frozen_cases": true,
        "one_measurement_per_case_per_arm": true,
        "luna_provider_reasoning": true,
        "jev_provider_model": true,
        "luna_cache_reads_and_writes_zero": true
      },
      "passed": true,
      "records": 364,
      "measured": 360,
      "raw_sha256": "501a97e6330c2c44126075afbc7a8f9cb0a39bfa5d07539833485a167112e87e",
      "billed_cost_usd": 0.01155007
    },
    "cache": {
      "checks": {
        "status_complete": true,
        "case_hash_matches": true,
        "source_hash_matches": true,
        "phase_counts_match": true,
        "consecutive_sequence_numbers": true,
        "unique_generation_ids": true,
        "http_200_and_valid": true,
        "exact_answer_schema": true,
        "probability_matches_answer": true,
        "logged_answer_matches_raw_response": true,
        "raw_cost_matches_record_and_finite": true,
        "billed_total_matches_metadata": true,
        "positive_finite_latencies": true,
        "targets_and_inputs_match_frozen_cases": true,
        "one_measurement_per_case_per_arm": true,
        "luna_provider_reasoning": true,
        "jev_provider_model": true,
        "all_cached_measured_have_reads": true,
        "uncached_reads_and_writes_zero": true,
        "actual_luna_payloads_differ_only_in_cache_breakpoint": true
      },
      "passed": true,
      "records": 73,
      "measured": 72,
      "raw_sha256": "b790eeb249847b3438ce79dfee24f74e9971dc649e43864b18ac353cac73b97c",
      "billed_cost_usd": 0.012438149
    },
    "existing_summary_metric_mismatches": [],
    "all_numeric_summaries_reconciled": true,
    "complete_arm_case_counts": {
      "main": {
        "Luna": 180,
        "Jev": 180
      },
      "cache": {
        "Luna uncached": 24,
        "Luna cached": 24,
        "Jev": 24
      }
    }
  },
  "main": {
    "policy": {
      "models": {
        "Jev": {
          "n": 60,
          "valid": 60,
          "median_latency_s": 0.24326154148729984,
          "p95_latency_s": 0.3312737014319282,
          "mean_latency_s": 0.25546772780217,
          "billed_cost_usd": 0.001255632,
          "usd_per_1000_requests": 0.020927200000000003,
          "exact_zero_or_one_predictions": 0,
          "correct": 60,
          "accuracy": 1.0,
          "accuracy_wilson95": [
            0.93982814785791,
            0.9999999999999998
          ],
          "brier": 0.008311666666666669,
          "log_loss": 0.06748769727937004,
          "incorrect_case_ids": []
        },
        "Luna": {
          "n": 60,
          "valid": 60,
          "median_latency_s": 1.1491409375012154,
          "p95_latency_s": 1.6466312964330425,
          "mean_latency_s": 1.1813501993514364,
          "billed_cost_usd": 0.0027441,
          "usd_per_1000_requests": 0.045735000000000005,
          "exact_zero_or_one_predictions": 60,
          "correct": 60,
          "accuracy": 1.0,
          "accuracy_wilson95": [
            0.93982814785791,
            0.9999999999999998
          ],
          "brier": 0.0,
          "log_loss": 9.99600361081321e-16,
          "incorrect_case_ids": []
        }
      },
      "paired": {
        "left": "Luna",
        "right": "Jev",
        "n_pairs": 60,
        "difference": "left minus right",
        "ratio": "left divided by right",
        "bootstrap_seed": 20261002,
        "metrics": {
          "brier_difference": {
            "estimate": -0.008311666666666669,
            "ci95": [
              -0.014858500000000002,
              -0.0037949583333333377
            ]
          },
          "mean_absolute_label_error_difference": {
            "estimate": -0.06250000000000001,
            "ci95": [
              -0.08050000000000002,
              -0.047333333333333345
            ]
          },
          "mean_latency_difference_s": {
            "estimate": 0.9258824715492665,
            "ci95": [
              0.8648824935581433,
              0.9946337015639558
            ]
          },
          "mean_cost_difference_usd": {
            "estimate": 2.48078e-05,
            "ci95": [
              2.4696865833333333e-05,
              2.49125e-05
            ]
          },
          "accuracy_difference": {
            "estimate": 0.0,
            "ci95": [
              0.0,
              0.0
            ]
          },
          "log_loss_difference": {
            "estimate": -0.06748769727936903,
            "ci95": [
              -0.0895977519743015,
              -0.04958625371843448
            ]
          },
          "ratio_of_median_latencies": {
            "estimate": 4.723890716450177,
            "ci95": [
              4.266531603721035,
              5.047853414741613
            ]
          },
          "median_of_paired_latency_ratios": {
            "estimate": 4.5148499480948425,
            "ci95": [
              4.1382223950328,
              4.789030734144223
            ]
          },
          "ratio_of_mean_costs": {
            "estimate": 2.1854333116709355,
            "ci95": [
              2.184486385094379,
              2.186404222462109
            ]
          }
        },
        "paired_correctness_counts": {
          "both_correct": 60,
          "left_only_correct": 0,
          "right_only_correct": 0,
          "both_incorrect": 0
        }
      },
      "family_descriptive_results": {
        "delivery": {
          "Jev": {
            "n": 20,
            "valid": 20,
            "median_latency_s": 0.23518043749209028,
            "p95_latency_s": 0.3075738895640825,
            "mean_latency_s": 0.24279662920162082,
            "billed_cost_usd": 0.000425964,
            "usd_per_1000_requests": 0.0212982,
            "exact_zero_or_one_predictions": 0,
            "correct": 20,
            "accuracy": 1.0,
            "accuracy_wilson95": [
              0.8388748419471806,
              1
            ],
            "brier": 0.011625000000000002,
            "log_loss": 0.07697036711430137,
            "incorrect_case_ids": []
          },
          "Luna": {
            "n": 20,
            "valid": 20,
            "median_latency_s": 1.1735261874855496,
            "p95_latency_s": 1.2876820315883382,
            "mean_latency_s": 1.133705837551679,
            "billed_cost_usd": 0.0009294,
            "usd_per_1000_requests": 0.04647,
            "exact_zero_or_one_predictions": 20,
            "correct": 20,
            "accuracy": 1.0,
            "accuracy_wilson95": [
              0.8388748419471806,
              1
            ],
            "brier": 0.0,
            "log_loss": 9.99600361081321e-16,
            "incorrect_case_ids": []
          }
        },
        "access": {
          "Jev": {
            "n": 20,
            "valid": 20,
            "median_latency_s": 0.24228670848242473,
            "p95_latency_s": 0.392510266667523,
            "mean_latency_s": 0.25893684795155425,
            "billed_cost_usd": 0.000421806,
            "usd_per_1000_requests": 0.0210903,
            "exact_zero_or_one_predictions": 0,
            "correct": 20,
            "accuracy": 1.0,
            "accuracy_wilson95": [
              0.8388748419471806,
              1
            ],
            "brier": 0.009595000000000001,
            "log_loss": 0.0730228734197132,
            "incorrect_case_ids": []
          },
          "Luna": {
            "n": 20,
            "valid": 20,
            "median_latency_s": 1.0688631250086473,
            "p95_latency_s": 1.5101126541529086,
            "mean_latency_s": 1.1539596479502507,
            "billed_cost_usd": 0.0009227,
            "usd_per_1000_requests": 0.046135,
            "exact_zero_or_one_predictions": 20,
            "correct": 20,
            "accuracy": 1.0,
            "accuracy_wilson95": [
              0.8388748419471806,
              1
            ],
            "brier": 0.0,
            "log_loss": 9.99600361081321e-16,
            "incorrect_case_ids": []
          }
        },
        "refund": {
          "Jev": {
            "n": 20,
            "valid": 20,
            "median_latency_s": 0.2587476255139336,
            "p95_latency_s": 0.3312737014319282,
            "mean_latency_s": 0.26466970625333486,
            "billed_cost_usd": 0.000407862,
            "usd_per_1000_requests": 0.0203931,
            "exact_zero_or_one_predictions": 0,
            "correct": 20,
            "accuracy": 1.0,
            "accuracy_wilson95": [
              0.8388748419471806,
              1
            ],
            "brier": 0.0037150000000000004,
            "log_loss": 0.052469851304095524,
            "incorrect_case_ids": []
          },
          "Luna": {
            "n": 20,
            "valid": 20,
            "median_latency_s": 1.177424396009883,
            "p95_latency_s": 1.9449935735450712,
            "mean_latency_s": 1.2563851125523797,
            "billed_cost_usd": 0.000892,
            "usd_per_1000_requests": 0.0446,
            "exact_zero_or_one_predictions": 20,
            "correct": 20,
            "accuracy": 1.0,
            "accuracy_wilson95": [
              0.8388748419471806,
              1
            ],
            "brier": 0.0,
            "log_loss": 9.99600361081321e-16,
            "incorrect_case_ids": []
          }
        }
      }
    },
    "boolq": {
      "models": {
        "Jev": {
          "n": 60,
          "valid": 60,
          "median_latency_s": 0.23620002050301991,
          "p95_latency_s": 0.37197747530590275,
          "mean_latency_s": 0.26132939653180076,
          "billed_cost_usd": 0.001190826,
          "usd_per_1000_requests": 0.019847100000000003,
          "exact_zero_or_one_predictions": 0,
          "correct": 57,
          "accuracy": 0.95,
          "accuracy_wilson95": [
            0.8629948352365147,
            0.9828504978356043
          ],
          "brier": 0.06835999999999999,
          "log_loss": 0.2482454306996313,
          "incorrect_case_ids": [
            "boolq-dev-1517",
            "boolq-dev-1842",
            "boolq-dev-0587"
          ]
        },
        "Luna": {
          "n": 60,
          "valid": 60,
          "median_latency_s": 1.1484179999970365,
          "p95_latency_s": 1.4701433298032498,
          "mean_latency_s": 1.1566619764028776,
          "billed_cost_usd": 0.002669,
          "usd_per_1000_requests": 0.04448333333333333,
          "exact_zero_or_one_predictions": 45,
          "correct": 49,
          "accuracy": 0.8166666666666667,
          "accuracy_wilson95": [
            0.7008024625353194,
            0.8944220311080237
          ],
          "brier": 0.16761500000000001,
          "log_loss": 3.6759119683872865,
          "incorrect_case_ids": [
            "boolq-dev-1447",
            "boolq-dev-2504",
            "boolq-dev-0448",
            "boolq-dev-1625",
            "boolq-dev-1517",
            "boolq-dev-2597",
            "boolq-dev-0852",
            "boolq-dev-2868",
            "boolq-dev-2665",
            "boolq-dev-1842",
            "boolq-dev-0587"
          ]
        }
      },
      "paired": {
        "left": "Luna",
        "right": "Jev",
        "n_pairs": 60,
        "difference": "left minus right",
        "ratio": "left divided by right",
        "bootstrap_seed": 20261003,
        "metrics": {
          "brier_difference": {
            "estimate": 0.09925500000000001,
            "ci95": [
              0.0335895,
              0.17571883333333327
            ]
          },
          "mean_absolute_label_error_difference": {
            "estimate": 0.04516666666666666,
            "ci95": [
              -0.024333333333333332,
              0.12366666666666666
            ]
          },
          "mean_latency_difference_s": {
            "estimate": 0.895332579871077,
            "ci95": [
              0.8495035622303112,
              0.941849156431296
            ]
          },
          "mean_cost_difference_usd": {
            "estimate": 2.4636233333333333e-05,
            "ci95": [
              2.3409253333333334e-05,
              2.6234952499999995e-05
            ]
          },
          "accuracy_difference": {
            "estimate": -0.13333333333333333,
            "ci95": [
              -0.23333333333333334,
              -0.05
            ]
          },
          "log_loss_difference": {
            "estimate": 3.4276665376876547,
            "ci95": [
              1.153912919291477,
              6.183977213123923
            ]
          },
          "ratio_of_median_latencies": {
            "estimate": 4.862057156266646,
            "ci95": [
              4.439819894127453,
              5.1176199255015336
            ]
          },
          "median_of_paired_latency_ratios": {
            "estimate": 4.39906621865231,
            "ci95": [
              4.244045901529168,
              4.9290712246743675
            ]
          },
          "ratio_of_mean_costs": {
            "estimate": 2.241301415991925,
            "ci95": [
              2.231369685908802,
              2.2524149376246383
            ]
          }
        },
        "paired_correctness_counts": {
          "both_correct": 49,
          "left_only_correct": 0,
          "right_only_correct": 8,
          "both_incorrect": 3
        }
      }
    },
    "probability": {
      "models": {
        "Jev": {
          "n": 60,
          "valid": 60,
          "median_latency_s": 0.237314312485978,
          "p95_latency_s": 0.31138571220799344,
          "mean_latency_s": 0.2471950327322702,
          "billed_cost_usd": 0.00111867,
          "usd_per_1000_requests": 0.0186445,
          "exact_zero_or_one_predictions": 0,
          "mae": 0.0644286219101022,
          "rmse": 0.09988414577509423,
          "excess_brier": 0.009976842577220275,
          "expected_brier": 0.15633600538274284
        },
        "Luna": {
          "n": 60,
          "valid": 60,
          "median_latency_s": 1.1426221465080744,
          "p95_latency_s": 1.5959135121971477,
          "mean_latency_s": 1.2395371187997322,
          "billed_cost_usd": 0.0024845,
          "usd_per_1000_requests": 0.04140833333333333,
          "exact_zero_or_one_predictions": 6,
          "mae": 0.03277039096880469,
          "rmse": 0.07423081008255741,
          "excess_brier": 0.005510213165512707,
          "expected_brier": 0.15186937597103528
        }
      },
      "paired": {
        "left": "Luna",
        "right": "Jev",
        "n_pairs": 60,
        "difference": "left minus right",
        "ratio": "left divided by right",
        "bootstrap_seed": 20261004,
        "metrics": {
          "excess_brier_difference": {
            "estimate": -0.004466629411707567,
            "ci95": [
              -0.011735129597998939,
              0.002056910350914978
            ]
          },
          "mae_difference": {
            "estimate": -0.0316582309412975,
            "ci95": [
              -0.05221159271243001,
              -0.011476891283283205
            ]
          },
          "mean_latency_difference_s": {
            "estimate": 0.992342086067462,
            "ci95": [
              0.917132350749198,
              1.079241541537558
            ]
          },
          "mean_cost_difference_usd": {
            "estimate": 2.2763833333333334e-05,
            "ci95": [
              2.2401362500000003e-05,
              2.3128505833333332e-05
            ]
          },
          "ratio_of_median_latencies": {
            "estimate": 4.814805034464946,
            "ci95": [
              4.480785852223475,
              5.36817426179022
            ]
          },
          "median_of_paired_latency_ratios": {
            "estimate": 4.750724711529367,
            "ci95": [
              4.48309361915998,
              5.076219965790033
            ]
          },
          "ratio_of_mean_costs": {
            "estimate": 2.220940938793389,
            "ci95": [
              2.2073309330627984,
              2.235412532162518
            ]
          }
        }
      },
      "family_descriptive_results": {
        "without_replacement": {
          "Jev": {
            "n": 20,
            "valid": 20,
            "median_latency_s": 0.24165856248873752,
            "p95_latency_s": 0.29094072920124747,
            "mean_latency_s": 0.25065561265219005,
            "billed_cost_usd": 0.000350154,
            "usd_per_1000_requests": 0.017507699999999998,
            "exact_zero_or_one_predictions": 0,
            "mae": 0.08215762313029003,
            "rmse": 0.11443929483405323,
            "excess_brier": 0.013096352202115363,
            "expected_brier": 0.13237820973353845
          },
          "Luna": {
            "n": 20,
            "valid": 20,
            "median_latency_s": 1.0988034585025162,
            "p95_latency_s": 1.4990901916447918,
            "mean_latency_s": 1.2055278500483837,
            "billed_cost_usd": 0.000792,
            "usd_per_1000_requests": 0.0396,
            "exact_zero_or_one_predictions": 3,
            "mae": 0.01884820341602098,
            "rmse": 0.06357040152690356,
            "excess_brier": 0.004041195950291742,
            "expected_brier": 0.12332305348171482
          }
        },
        "weighted_mixture": {
          "Jev": {
            "n": 20,
            "valid": 20,
            "median_latency_s": 0.23649129198747687,
            "p95_latency_s": 0.3113857122079935,
            "mean_latency_s": 0.24089466054720104,
            "billed_cost_usd": 0.00040005,
            "usd_per_1000_requests": 0.0200025,
            "exact_zero_or_one_predictions": 0,
            "mae": 0.09071157593334987,
            "rmse": 0.1277612051036834,
            "excess_brier": 0.016322925529545458,
            "expected_brier": 0.20021147308135673
          },
          "Luna": {
            "n": 20,
            "valid": 20,
            "median_latency_s": 1.188938417006284,
            "p95_latency_s": 1.9231115711547324,
            "mean_latency_s": 1.2903996457986069,
            "billed_cost_usd": 0.000888,
            "usd_per_1000_requests": 0.0444,
            "exact_zero_or_one_predictions": 0,
            "mae": 0.07946296948705976,
            "rmse": 0.11175617900700784,
            "excess_brier": 0.012489443546246379,
            "expected_brier": 0.19637799109805767
          }
        },
        "conditional_table": {
          "Jev": {
            "n": 20,
            "valid": 20,
            "median_latency_s": 0.24253977098851465,
            "p95_latency_s": 0.2937989497047966,
            "mean_latency_s": 0.25003482499741947,
            "billed_cost_usd": 0.000368466,
            "usd_per_1000_requests": 0.0184233,
            "exact_zero_or_one_predictions": 0,
            "mae": 0.020416666666666666,
            "rmse": 0.022610838109190023,
            "excess_brier": 0.0005112499999999999,
            "expected_brier": 0.13641833333333336
          },
          "Luna": {
            "n": 20,
            "valid": 20,
            "median_latency_s": 1.2159329165006056,
            "p95_latency_s": 1.5020713753503514,
            "mean_latency_s": 1.2226838605522061,
            "billed_cost_usd": 0.0008045,
            "usd_per_1000_requests": 0.040225000000000004,
            "exact_zero_or_one_predictions": 3,
            "mae": 3.33333360913457e-12,
            "rmse": 1.0540926406054687e-11,
            "excess_brier": 1.1111112949786098e-22,
            "expected_brier": 0.13590708333333335
          }
        }
      }
    }
  },
  "cache": {
    "models": {
      "Luna cached": {
        "n": 24,
        "valid": 24,
        "median_latency_s": 0.9179664584953571,
        "p95_latency_s": 1.1803516982632571,
        "mean_latency_s": 0.9428539184179195,
        "billed_cost_usd": 0.00171132,
        "usd_per_1000_requests": 0.071305,
        "exact_zero_or_one_predictions": 24,
        "correct": 21,
        "accuracy": 0.875,
        "accuracy_wilson95": [
          0.6899611872949993,
          0.9565566591987655
        ],
        "brier": 0.125,
        "log_loss": 4.317413682483019,
        "incorrect_case_ids": [
          "cache-member-supervised-expired",
          "cache-sensor-outdoor-weather",
          "cache-portable-battery-low"
        ]
      },
      "Luna uncached": {
        "n": 24,
        "valid": 24,
        "median_latency_s": 1.0602054579940159,
        "p95_latency_s": 1.291354949955712,
        "mean_latency_s": 1.136900599000607,
        "billed_cost_usd": 0.0071802,
        "usd_per_1000_requests": 0.29917499999999997,
        "exact_zero_or_one_predictions": 24,
        "correct": 21,
        "accuracy": 0.875,
        "accuracy_wilson95": [
          0.6899611872949993,
          0.9565566591987655
        ],
        "brier": 0.125,
        "log_loss": 4.317413682483019,
        "incorrect_case_ids": [
          "cache-after-hours-permitted",
          "cache-sensor-outdoor-weather",
          "cache-portable-battery-low"
        ]
      },
      "Jev": {
        "n": 24,
        "valid": 24,
        "median_latency_s": 0.2544942910026293,
        "p95_latency_s": 0.33727485434792465,
        "mean_latency_s": 0.2704224321253908,
        "billed_cost_usd": 0.003184104,
        "usd_per_1000_requests": 0.132671,
        "exact_zero_or_one_predictions": 0,
        "correct": 20,
        "accuracy": 0.8333333333333334,
        "accuracy_wilson95": [
          0.6414692935030095,
          0.9332132367136705
        ],
        "brier": 0.08786666666666666,
        "log_loss": 0.28680211062673244,
        "incorrect_case_ids": [
          "cache-sensor-outdoor-weather",
          "cache-after-hours-unsigned",
          "cache-trainee-duration-limit",
          "cache-portable-battery-low"
        ]
      }
    },
    "paired": {
      "Luna cached vs Luna uncached": {
        "left": "Luna cached",
        "right": "Luna uncached",
        "n_pairs": 24,
        "difference": "left minus right",
        "ratio": "left divided by right",
        "bootstrap_seed": 20261012,
        "metrics": {
          "brier_difference": {
            "estimate": 0.0,
            "ci95": [
              -0.125,
              0.125
            ]
          },
          "mean_absolute_label_error_difference": {
            "estimate": 0.0,
            "ci95": [
              -0.125,
              0.125
            ]
          },
          "mean_latency_difference_s": {
            "estimate": -0.19404668058268726,
            "ci95": [
              -0.3724629672147785,
              -0.07326244343109173
            ]
          },
          "mean_cost_difference_usd": {
            "estimate": -0.00022787000000000003,
            "ci95": [
              -0.00022787000000000003,
              -0.00022787000000000003
            ]
          },
          "accuracy_difference": {
            "estimate": 0.0,
            "ci95": [
              -0.125,
              0.125
            ]
          },
          "log_loss_difference": {
            "estimate": 0.0,
            "ci95": [
              -4.3173470493638355,
              4.3173470493638355
            ]
          },
          "ratio_of_median_latencies": {
            "estimate": 0.8658382689636543,
            "ci95": [
              0.8002578901736012,
              0.9627587174487656
            ]
          },
          "median_of_paired_latency_ratios": {
            "estimate": 0.9125562576625701,
            "ci95": [
              0.7795041957054406,
              0.9593429075486408
            ]
          },
          "ratio_of_mean_costs": {
            "estimate": 0.2383387649369098,
            "ci95": [
              0.2382645030991016,
              0.23841301229650877
            ]
          },
          "setup_inclusive_cost_ratio_fixed_observed_prime": {
            "estimate": 0.28882830561822787,
            "ci95": [
              0.2887589665018455,
              0.28889763121614276
            ]
          }
        },
        "paired_correctness_counts": {
          "both_correct": 20,
          "left_only_correct": 1,
          "right_only_correct": 1,
          "both_incorrect": 2
        }
      },
      "Luna cached vs Jev": {
        "left": "Luna cached",
        "right": "Jev",
        "n_pairs": 24,
        "difference": "left minus right",
        "ratio": "left divided by right",
        "bootstrap_seed": 20261013,
        "metrics": {
          "brier_difference": {
            "estimate": 0.03713333333333333,
            "ci95": [
              -0.06411041666666667,
              0.1536670833333333
            ]
          },
          "mean_absolute_label_error_difference": {
            "estimate": -0.08166666666666668,
            "ci95": [
              -0.18625,
              0.02666666666666666
            ]
          },
          "mean_latency_difference_s": {
            "estimate": 0.6724314862925288,
            "ci95": [
              0.6306404519296241,
              0.7208375596278225
            ]
          },
          "mean_cost_difference_usd": {
            "estimate": -6.136600000000001e-05,
            "ci95": [
              -6.138725e-05,
              -6.134341666666668e-05
            ]
          },
          "accuracy_difference": {
            "estimate": 0.041666666666666664,
            "ci95": [
              -0.08333333333333333,
              0.16666666666666666
            ]
          },
          "log_loss_difference": {
            "estimate": 4.030611571856286,
            "ci95": [
              -0.18489103719648087,
              8.363371197754898
            ]
          },
          "ratio_of_median_latencies": {
            "estimate": 3.6070218112903487,
            "ci95": [
              3.402496772316425,
              3.7383018313273846
            ]
          },
          "median_of_paired_latency_ratios": {
            "estimate": 3.617968437782495,
            "ci95": [
              3.4348607259371517,
              3.6970101768802643
            ]
          },
          "ratio_of_mean_costs": {
            "estimate": 0.5374573192332913,
            "ci95": [
              0.5372911169744943,
              0.5376386635537441
            ]
          },
          "setup_inclusive_cost_ratio_fixed_observed_prime": {
            "estimate": 0.6513119546346475,
            "ci95": [
              0.6511492855792343,
              0.6514888293051501
            ]
          }
        },
        "paired_correctness_counts": {
          "both_correct": 19,
          "left_only_correct": 2,
          "right_only_correct": 1,
          "both_incorrect": 2
        }
      },
      "Luna uncached vs Jev": {
        "left": "Luna uncached",
        "right": "Jev",
        "n_pairs": 24,
        "difference": "left minus right",
        "ratio": "left divided by right",
        "bootstrap_seed": 20261014,
        "metrics": {
          "brier_difference": {
            "estimate": 0.03713333333333333,
            "ci95": [
              -0.06677614583333333,
              0.16213062499999986
            ]
          },
          "mean_absolute_label_error_difference": {
            "estimate": -0.08166666666666668,
            "ci95": [
              -0.19208333333333336,
              0.040833333333333326
            ]
          },
          "mean_latency_difference_s": {
            "estimate": 0.8664781668752161,
            "ci95": [
              0.755344436835882,
              1.037138564075758
            ]
          },
          "mean_cost_difference_usd": {
            "estimate": 0.00016650400000000002,
            "ci95": [
              0.0001664825,
              0.00016652683333333337
            ]
          },
          "accuracy_difference": {
            "estimate": 0.041666666666666664,
            "ci95": [
              -0.08333333333333333,
              0.16666666666666666
            ]
          },
          "log_loss_difference": {
            "estimate": 4.030611571856286,
            "ci95": [
              -0.18605978817093258,
              9.448219672079865
            ]
          },
          "ratio_of_median_latencies": {
            "estimate": 4.165930221134361,
            "ci95": [
              3.7397323426328377,
              4.46620313518592
            ]
          },
          "median_of_paired_latency_ratios": {
            "estimate": 3.918573737604335,
            "ci95": [
              3.6925993959134398,
              4.406468553807654
            ]
          },
          "ratio_of_mean_costs": {
            "estimate": 2.255014283453053,
            "ci95": [
              2.254743282717801,
              2.2552820186159237
            ]
          }
        },
        "paired_correctness_counts": {
          "both_correct": 19,
          "left_only_correct": 2,
          "right_only_correct": 1,
          "both_incorrect": 2
        }
      }
    },
    "priming": {
      "cost_usd": 0.000362525,
      "latency_s": 1.2044517089962028,
      "read_tokens": 0,
      "written_tokens": 2533,
      "inference_limit": "Single observed priming request; setup-inclusive interval treats its measured bill as fixed and does not estimate uncertainty in future priming latency/cost."
    },
    "cache_counts": {
      "cached_arm_has_observed_hits": true,
      "uncached_control_verified": true,
      "controlled_cache_comparison_verified": true,
      "status": "observed cache hits and clean uncached control",
      "prime_cached_tokens": 0,
      "prime_cache_write_tokens": 2533,
      "prime_cost_usd": 0.000362525,
      "prime_latency_s": 1.2044517089962028,
      "prime_was_observed_cold": true,
      "setup_inclusive_cached_cost_usd": 0.002073845,
      "setup_inclusive_cost_per_measured_case_usd": 8.641020833333333e-05
    }
  },
  "billing": {
    "main_measured_usd": 0.011462728,
    "main_warmups_usd": 8.7342e-05,
    "setup_probes_usd": 8.7342e-05,
    "cache_measured_usd": 0.012075624,
    "cache_prime_usd": 0.000362525,
    "all_observed_requests": 441,
    "all_observed_billed_usd": 0.024075561
  },
  "boolq_posthoc_exclusion_sensitivity": {
    "label": "Exclude only the one pre-existing audit flag for an unambiguous passage-label contradiction; keep all other labels and cases unchanged",
    "excluded_case_id": "boolq-dev-1842",
    "primary_label": 1,
    "passage_supported_label": 0,
    "predictions": {
      "Luna": 0.0,
      "Jev": 0.03
    },
    "models": {
      "Jev": {
        "n": 59,
        "valid": 59,
        "median_latency_s": 0.2360435409937054,
        "p95_latency_s": 0.3724394086049869,
        "mean_latency_s": 0.2597932514050873,
        "billed_cost_usd": 0.001172598,
        "usd_per_1000_requests": 0.019874542372881356,
        "exact_zero_or_one_predictions": 0,
        "correct": 57,
        "accuracy": 0.9661016949152542,
        "accuracy_wilson95": [
          0.8845642924988925,
          0.9906540909083387
        ],
        "brier": 0.053571186440677966,
        "log_loss": 0.1930197956721677,
        "incorrect_case_ids": [
          "boolq-dev-1517",
          "boolq-dev-0587"
        ]
      },
      "Luna": {
        "n": 59,
        "valid": 59,
        "median_latency_s": 1.1432648339832667,
        "p95_latency_s": 1.4705759506032337,
        "mean_latency_s": 1.15482401625713,
        "billed_cost_usd": 0.0026284,
        "usd_per_1000_requests": 0.044549152542372884,
        "exact_zero_or_one_predictions": 44,
        "correct": 49,
        "accuracy": 0.8305084745762712,
        "accuracy_wilson95": [
          0.71536779263512,
          0.9052416065081895
        ],
        "brier": 0.15350677966101697,
        "log_loss": 3.152812571327568,
        "incorrect_case_ids": [
          "boolq-dev-1447",
          "boolq-dev-2504",
          "boolq-dev-0448",
          "boolq-dev-1625",
          "boolq-dev-1517",
          "boolq-dev-2597",
          "boolq-dev-0852",
          "boolq-dev-2868",
          "boolq-dev-2665",
          "boolq-dev-0587"
        ]
      }
    },
    "paired": {
      "left": "Luna",
      "right": "Jev",
      "n_pairs": 59,
      "difference": "left minus right",
      "ratio": "left divided by right",
      "bootstrap_seed": 20261022,
      "metrics": {
        "brier_difference": {
          "estimate": 0.09993559322033899,
          "ci95": [
            0.033645466101694915,
            0.17669563559322013
          ]
        },
        "mean_absolute_label_error_difference": {
          "estimate": 0.04542372881355932,
          "ci95": [
            -0.025427966101694923,
            0.12373305084745725
          ]
        },
        "mean_latency_difference_s": {
          "estimate": 0.8950307648520427,
          "ci95": [
            0.8485650032077363,
            0.9422056648823516
          ]
        },
        "mean_cost_difference_usd": {
          "estimate": 2.4674610169491526e-05,
          "ci95": [
            2.3425851694915254e-05,
            2.623991864406779e-05
          ]
        },
        "accuracy_difference": {
          "estimate": -0.13559322033898305,
          "ci95": [
            -0.23728813559322035,
            -0.05084745762711865
          ]
        },
        "log_loss_difference": {
          "estimate": 2.9597927756553997,
          "ci95": [
            0.7258945316229899,
            5.670080571804208
          ]
        },
        "ratio_of_median_latencies": {
          "estimate": 4.84344892120456,
          "ci95": [
            4.439952214167102,
            5.07630224189156
          ]
        },
        "median_of_paired_latency_ratios": {
          "estimate": 4.400885207005879,
          "ci95": [
            4.269253980950404,
            4.9290712246743675
          ]
        },
        "ratio_of_mean_costs": {
          "estimate": 2.241518406137483,
          "ci95": [
            2.23132303121675,
            2.2525479038346683
          ]
        }
      },
      "paired_correctness_counts": {
        "both_correct": 49,
        "left_only_correct": 0,
        "right_only_correct": 8,
        "both_incorrect": 2
      }
    }
  },
  "data_limitations": {
    "boolq_unique_passages": 59,
    "boolq_repeated_passage_case_groups": [
      [
        "boolq-dev-0659",
        "boolq-dev-0383"
      ]
    ],
    "boolq_audit_status_counts": {
      "supported": 38,
      "ambiguous": 15,
      "time_or_scope": 6,
      "contradiction": 1
    },
    "boolq_scope": "Agreement with frozen published human labels; not verified truth and not a current-law/current-fact evaluation.",
    "synthetic_main_design": "20 cases per family; three policy families with balanced labels and three probability families. Policy boundary/exception anchors are deliberately selected, not an IID random natural workload sample.",
    "cache_design": "24 hand-designed cases sharing a single synthetic rulebook; all 3 arms tested once/case; 24/24 cached calls had reads; uncached control reads and writes all zero.",
    "schema_equivalence": "The returned answer object matches exactly. API envelopes, tokenizers, backend processing and required Luna formatting instruction differ."
  }
}
