{
  "schema_version": 1,
  "updated": "2026-09-28",
  "rows": [
    {
      "section": "general",
      "cohort": "core",
      "model": "JEV 1.13",
      "correct": 1715,
      "total": 1800,
      "packets": 180,
      "questions_per_request": 10,
      "median_seconds": 0.2683277000032831,
      "cost_per_1000_packets": 0.11011606666666669,
      "deployment": "API",
      "false_positives": null,
      "false_negatives": null,
      "routes_correct": 180,
      "wrong_clears": 0,
      "unnecessary_holds": 0,
      "gpu_gib": null,
      "date": "2026-09-26"
    },
    {
      "section": "binary",
      "cohort": "core",
      "model": "JEV 1.13",
      "correct": 513,
      "total": 540,
      "packets": 180,
      "questions_per_request": 10,
      "median_seconds": 0.2683277000032831,
      "cost_per_1000_packets": 0.11011606666666669,
      "deployment": "API",
      "false_positives": 27,
      "false_negatives": 0,
      "routes_correct": null,
      "wrong_clears": null,
      "unnecessary_holds": null,
      "gpu_gib": null,
      "date": "2026-09-26"
    },
    {
      "section": "general",
      "cohort": "core",
      "model": "KEV-4B",
      "correct": 1575,
      "total": 1800,
      "packets": 180,
      "questions_per_request": 10,
      "median_seconds": 0.759520750012598,
      "cost_per_1000_packets": 0.08427860000000001,
      "deployment": "API",
      "false_positives": null,
      "false_negatives": null,
      "routes_correct": 144,
      "wrong_clears": 0,
      "unnecessary_holds": 36,
      "gpu_gib": null,
      "date": "2026-09-26"
    },
    {
      "section": "binary",
      "cohort": "core",
      "model": "KEV-4B",
      "correct": 453,
      "total": 540,
      "packets": 180,
      "questions_per_request": 10,
      "median_seconds": 0.759520750012598,
      "cost_per_1000_packets": 0.08427860000000001,
      "deployment": "API",
      "false_positives": 87,
      "false_negatives": 0,
      "routes_correct": null,
      "wrong_clears": null,
      "unnecessary_holds": null,
      "gpu_gib": null,
      "date": "2026-09-26"
    },
    {
      "section": "binary",
      "cohort": "core",
      "model": "Span-01",
      "correct": 497,
      "total": 540,
      "packets": 180,
      "questions_per_request": 3,
      "median_seconds": 0.49353680000058375,
      "cost_per_1000_packets": 0.04076877777777778,
      "deployment": "API",
      "false_positives": 43,
      "false_negatives": 0,
      "routes_correct": null,
      "wrong_clears": null,
      "unnecessary_holds": null,
      "gpu_gib": null,
      "date": "2026-09-27"
    },
    {
      "section": "binary",
      "cohort": "core",
      "model": "Span-01 Lite",
      "correct": 350,
      "total": 540,
      "packets": 180,
      "questions_per_request": 3,
      "median_seconds": 0.507799699989846,
      "cost_per_1000_packets": 0.0,
      "deployment": "API",
      "false_positives": 141,
      "false_negatives": 49,
      "routes_correct": null,
      "wrong_clears": null,
      "unnecessary_holds": null,
      "gpu_gib": null,
      "date": "2026-09-27"
    },
    {
      "section": "general",
      "cohort": "smoke",
      "model": "JEV 1.13",
      "correct": 78,
      "total": 80,
      "packets": 8,
      "questions_per_request": 10,
      "median_seconds": 0.2701378499768907,
      "cost_per_1000_packets": 0.107688,
      "deployment": "API",
      "false_positives": null,
      "false_negatives": null,
      "routes_correct": 8,
      "wrong_clears": 0,
      "unnecessary_holds": 0,
      "gpu_gib": null,
      "date": "2026-09-26"
    },
    {
      "section": "binary",
      "cohort": "smoke",
      "model": "JEV 1.13",
      "correct": 23,
      "total": 24,
      "packets": 8,
      "questions_per_request": 10,
      "median_seconds": 0.2701378499768907,
      "cost_per_1000_packets": 0.107688,
      "deployment": "API",
      "false_positives": 1,
      "false_negatives": 0,
      "routes_correct": null,
      "wrong_clears": null,
      "unnecessary_holds": null,
      "gpu_gib": null,
      "date": "2026-09-26"
    },
    {
      "section": "general",
      "cohort": "smoke",
      "model": "KEV-4B",
      "correct": 72,
      "total": 80,
      "packets": 8,
      "questions_per_request": 10,
      "median_seconds": 0.7503021499869647,
      "cost_per_1000_packets": 0.08186325,
      "deployment": "API",
      "false_positives": null,
      "false_negatives": null,
      "routes_correct": 8,
      "wrong_clears": 0,
      "unnecessary_holds": 0,
      "gpu_gib": null,
      "date": "2026-09-26"
    },
    {
      "section": "binary",
      "cohort": "smoke",
      "model": "KEV-4B",
      "correct": 22,
      "total": 24,
      "packets": 8,
      "questions_per_request": 10,
      "median_seconds": 0.7503021499869647,
      "cost_per_1000_packets": 0.08186325,
      "deployment": "API",
      "false_positives": 2,
      "false_negatives": 0,
      "routes_correct": null,
      "wrong_clears": null,
      "unnecessary_holds": null,
      "gpu_gib": null,
      "date": "2026-09-26"
    },
    {
      "section": "general",
      "cohort": "smoke",
      "model": "KEV-4B \u00b7 BF16",
      "correct": 72,
      "total": 80,
      "packets": 8,
      "questions_per_request": 10,
      "median_seconds": 5.865846500004409,
      "cost_per_1000_packets": null,
      "deployment": "Local GPU",
      "false_positives": null,
      "false_negatives": null,
      "routes_correct": 8,
      "wrong_clears": 0,
      "unnecessary_holds": 0,
      "gpu_gib": 10.198810577392578,
      "date": "2026-09-27"
    },
    {
      "section": "binary",
      "cohort": "smoke",
      "model": "KEV-4B \u00b7 BF16",
      "correct": 22,
      "total": 24,
      "packets": 8,
      "questions_per_request": 10,
      "median_seconds": 5.865846500004409,
      "cost_per_1000_packets": null,
      "deployment": "Local GPU",
      "false_positives": 2,
      "false_negatives": 0,
      "routes_correct": null,
      "wrong_clears": null,
      "unnecessary_holds": null,
      "gpu_gib": 10.198810577392578,
      "date": "2026-09-27"
    },
    {
      "section": "general",
      "cohort": "smoke",
      "model": "KEV-0.8B \u00b7 BF16",
      "correct": 41,
      "total": 80,
      "packets": 8,
      "questions_per_request": 10,
      "median_seconds": 0.9796066999842878,
      "cost_per_1000_packets": null,
      "deployment": "Local GPU",
      "false_positives": null,
      "false_negatives": null,
      "routes_correct": 3,
      "wrong_clears": 0,
      "unnecessary_holds": 3,
      "gpu_gib": 2.4486145973205566,
      "date": "2026-09-27"
    },
    {
      "section": "binary",
      "cohort": "smoke",
      "model": "KEV-0.8B \u00b7 BF16",
      "correct": 17,
      "total": 24,
      "packets": 8,
      "questions_per_request": 10,
      "median_seconds": 0.9796066999842878,
      "cost_per_1000_packets": null,
      "deployment": "Local GPU",
      "false_positives": 7,
      "false_negatives": 0,
      "routes_correct": null,
      "wrong_clears": null,
      "unnecessary_holds": null,
      "gpu_gib": 2.4486145973205566,
      "date": "2026-09-27"
    },
    {
      "section": "general",
      "cohort": "smoke",
      "model": "KEV-0.8B \u00b7 FP32",
      "correct": 40,
      "total": 80,
      "packets": 8,
      "questions_per_request": 10,
      "median_seconds": 0.6180222000111826,
      "cost_per_1000_packets": null,
      "deployment": "Local GPU",
      "false_positives": null,
      "false_negatives": null,
      "routes_correct": 2,
      "wrong_clears": 0,
      "unnecessary_holds": 4,
      "gpu_gib": 3.931704521179199,
      "date": "2026-09-27"
    },
    {
      "section": "binary",
      "cohort": "smoke",
      "model": "KEV-0.8B \u00b7 FP32",
      "correct": 16,
      "total": 24,
      "packets": 8,
      "questions_per_request": 10,
      "median_seconds": 0.6180222000111826,
      "cost_per_1000_packets": null,
      "deployment": "Local GPU",
      "false_positives": 8,
      "false_negatives": 0,
      "routes_correct": null,
      "wrong_clears": null,
      "unnecessary_holds": null,
      "gpu_gib": 3.931704521179199,
      "date": "2026-09-27"
    },
    {
      "section": "general",
      "cohort": "smoke",
      "model": "Julia-1",
      "correct": 33,
      "total": 80,
      "packets": 8,
      "questions_per_request": 10,
      "median_seconds": 0.7073948500037659,
      "cost_per_1000_packets": null,
      "deployment": "Local GPU",
      "false_positives": null,
      "false_negatives": null,
      "routes_correct": 3,
      "wrong_clears": 0,
      "unnecessary_holds": 4,
      "gpu_gib": 3.038177490234375,
      "date": "2026-09-28"
    },
    {
      "section": "binary",
      "cohort": "smoke",
      "model": "Julia-1",
      "correct": 13,
      "total": 24,
      "packets": 8,
      "questions_per_request": 10,
      "median_seconds": 0.7073948500037659,
      "cost_per_1000_packets": null,
      "deployment": "Local GPU",
      "false_positives": 9,
      "false_negatives": 2,
      "routes_correct": null,
      "wrong_clears": null,
      "unnecessary_holds": null,
      "gpu_gib": 3.038177490234375,
      "date": "2026-09-28"
    },
    {
      "section": "general",
      "cohort": "smoke",
      "model": "Decider-2B v11",
      "correct": 61,
      "total": 80,
      "packets": 8,
      "questions_per_request": 10,
      "median_seconds": 13.136242000007769,
      "cost_per_1000_packets": null,
      "deployment": "Local GPU",
      "false_positives": null,
      "false_negatives": null,
      "routes_correct": 6,
      "wrong_clears": 0,
      "unnecessary_holds": 2,
      "gpu_gib": 4.3595170974731445,
      "date": "2026-09-28"
    },
    {
      "section": "binary",
      "cohort": "smoke",
      "model": "Decider-2B v11",
      "correct": 23,
      "total": 24,
      "packets": 8,
      "questions_per_request": 10,
      "median_seconds": 13.136242000007769,
      "cost_per_1000_packets": null,
      "deployment": "Local GPU",
      "false_positives": 1,
      "false_negatives": 0,
      "routes_correct": null,
      "wrong_clears": null,
      "unnecessary_holds": null,
      "gpu_gib": 4.3595170974731445,
      "date": "2026-09-28"
    },
    {
      "section": "general",
      "cohort": "smoke",
      "model": "Solar Decide",
      "correct": 60,
      "total": 80,
      "packets": 8,
      "questions_per_request": 10,
      "median_seconds": 1.3107373999955598,
      "cost_per_1000_packets": 0.80321875,
      "deployment": "API",
      "false_positives": null,
      "false_negatives": null,
      "routes_correct": 5,
      "wrong_clears": 0,
      "unnecessary_holds": 3,
      "gpu_gib": null,
      "date": "2026-09-28"
    },
    {
      "section": "binary",
      "cohort": "smoke",
      "model": "Solar Decide",
      "correct": 24,
      "total": 24,
      "packets": 8,
      "questions_per_request": 10,
      "median_seconds": 1.3107373999955598,
      "cost_per_1000_packets": 0.80321875,
      "deployment": "API",
      "false_positives": 0,
      "false_negatives": 0,
      "routes_correct": null,
      "wrong_clears": null,
      "unnecessary_holds": null,
      "gpu_gib": null,
      "date": "2026-09-28"
    }
  ],
  "diagnostics": [
    {
      "model": "Laya \u00b7 typed-decisions",
      "general_correct": 46,
      "general_total": 80,
      "binary_correct": 16,
      "binary_total": 24,
      "note": "Adapted-input diagnostic: source-focused evidence, bound questions and compact policy. Best observed variant after exploratory changes, not a matched-input leaderboard result. No task-specific fine-tuning."
    }
  ],
  "sources": [
    {
      "artifact": "decision_decider_smoke_v01/ledger.jsonl",
      "sha256": "f140e873a8e4563ef7784b927a876b96c316df6bc70823b7f1cdd416f75fe3ea"
    },
    {
      "artifact": "decision_jk_lean_v01/ledger.jsonl",
      "sha256": "980df4155e519fd41f6d1fc84f81d54b990e25987c7e546cc2c399f5b60f0e53"
    },
    {
      "artifact": "decision_jk_lean_v01/reuse.json",
      "sha256": "9f463d037a07e5cbd1f6b6b7fd7f90e88e596dd807af2ba8061b2791e32316e0"
    },
    {
      "artifact": "decision_jk_lean_v01/summary.json",
      "sha256": "74e9dbd5ba6ae75afe18d9ae4beef4e2810c28301abc17279f16909b8461667e"
    },
    {
      "artifact": "decision_julia_smoke_v01/ledger.jsonl",
      "sha256": "8a5e33413ffb208a43925db86bb5eb32ed0a5ad9e26c51df9ba4c7abd0957668"
    },
    {
      "artifact": "decision_kev_08b_fp32_v01/ledger.jsonl",
      "sha256": "3d11119c184d0d4612ef8464dd3881bf899ffcb8719c3ef2904623bbf0199744"
    },
    {
      "artifact": "decision_kev_08b_smoke_v01/ledger.jsonl",
      "sha256": "6749e8ddbb82f929619ce925996089653b7b7b05954c3b2b60a4c51764d9c17f"
    },
    {
      "artifact": "decision_kev_idle_bf16_v01/ledger.jsonl",
      "sha256": "d3f3e3d8d70950cd8409d7bee4e328ad054dd860d553496a496446f1b9ca354c"
    },
    {
      "artifact": "decision_laya_compact_v01/summary.json",
      "sha256": "b91e1e54561e0d59c147e67dcace67adb7482234f7f3463aa452ed1ff82377ac"
    },
    {
      "artifact": "decision_solar_smoke_v01/answers.private.json",
      "sha256": "05899ccebe88fcbe12a6cb99f1efcab53917909c7fe348cf50123c8f243d7df6"
    },
    {
      "artifact": "decision_solar_smoke_v01/summary.json",
      "sha256": "a2c273dde14e7ccc8cac83fecfe057a69340db7cbdbea7e19c9441edb542fd68"
    },
    {
      "artifact": "decision_span_v01/summary.json",
      "sha256": "62c8fe5a4e86062d7a271b13480234863318539a6d02fb24fb3dd19c38d7c859"
    }
  ]
}