{
  "schema": "mp-capability-tournament-summary.v1",
  "status": "completed",
  "run_date": "2026-08-11",
  "study_class": "constructed_conformance",
  "manifest_hash": "sha256:e65d843669b1a0ead2a468ed8f05a44f3d74cf6e8184c05d2f697e427a8ec4ff",
  "unit_of_replication": {
    "cases": 8,
    "within_case_dispositions": 16,
    "total_scored_dispositions": 128,
    "independence_warning": "The 128 dispositions are not 128 independent trials."
  },
  "condition_boundary": {
    "A": "complete supplied lineage; reasoning only",
    "B": "the same complete supplied lineage; tools available",
    "C": "standalone deterministic canonical root vote; not the same model plus Minority Prophet",
    "standard": "fixed conventional methods"
  },
  "supported_claim": "On the frozen eight-case complete-lineage packet, the pinned canonical implementation conformed to all 128 within-case reference dispositions.",
  "unsupported_estimands": [
    "provenance gain",
    "Minority Prophet gain",
    "total epistemic gain",
    "H1",
    "H2",
    "H3",
    "stable model ranking",
    "production latency or cost"
  ],
  "threats_to_validity": [
    "reference is defined by the same distinct-origin invariant exercised by canonical C",
    "one clean replicate per model and lane",
    "within-case dispositions are correlated",
    "complete truthful lineage is supplied by construction",
    "no missing, forged, stale, or adversarial provenance is tested",
    "hosted model aliases may change",
    "CLI wall time and cost telemetry are not controlled serving measurements",
    "public packets are unsuitable as a future hidden evaluation set",
    "review and execution remain within one control domain unless externally verified"
  ],
  "results": [
    {"contestant":"Minority Prophet","lane":"C","protocol_correct":128,"exact_cases":8,"invalid_trials":0,"execution_ms":18.7,"tool_events":0,"input_tokens":0,"output_tokens":0,"cost_estimate_usd":0},
    {"contestant":"GPT-5.6 Terra","lane":"A","protocol_correct":116,"exact_cases":5,"invalid_trials":0,"execution_ms":365015,"tool_events":0,"input_tokens":300475,"output_tokens":16746,"cost_estimate_usd":0.8900575},
    {"contestant":"Claude Opus 5","lane":"A","protocol_correct":106,"exact_cases":6,"invalid_trials":0,"execution_ms":478018,"tool_events":0,"input_tokens":326569,"output_tokens":41681,"cost_estimate_usd":3.2534465},
    {"contestant":"GPT-5.6 Sol","lane":"A","protocol_correct":102,"exact_cases":5,"invalid_trials":0,"execution_ms":512249,"tool_events":0,"input_tokens":300694,"output_tokens":25952,"cost_estimate_usd":2.237102},
    {"contestant":"Cluster vote","lane":"STANDARD","protocol_correct":96,"exact_cases":6,"invalid_trials":0,"execution_ms":9.5,"tool_events":0,"input_tokens":0,"output_tokens":0,"cost_estimate_usd":0},
    {"contestant":"GPT-5.6 Sol","lane":"B","protocol_correct":69,"exact_cases":4,"invalid_trials":0,"execution_ms":351070,"tool_events":19,"input_tokens":1101609,"output_tokens":12559,"cost_estimate_usd":2.418447},
    {"contestant":"GPT-5.6 Terra","lane":"B","protocol_correct":68,"exact_cases":4,"invalid_trials":0,"execution_ms":245782,"tool_events":18,"input_tokens":1097326,"output_tokens":9151,"cost_estimate_usd":1.021252},
    {"contestant":"GPT-5.6 Luna","lane":"B","protocol_correct":35,"exact_cases":2,"invalid_trials":0,"execution_ms":496317,"tool_events":41,"input_tokens":1933985,"output_tokens":20412,"cost_estimate_usd":0.8089608},
    {"contestant":"Claude Sonnet 5","lane":"B","protocol_correct":32,"raw_correct":74,"exact_cases":2,"raw_exact_cases":4,"invalid_trials":5,"execution_ms":1609078,"tool_events":27,"input_tokens":2431595,"output_tokens":199228,"cost_estimate_usd":6.4535862},
    {"contestant":"Claude Sonnet 5","lane":"A","protocol_correct":23,"exact_cases":0,"invalid_trials":0,"execution_ms":809652,"tool_events":0,"input_tokens":639869,"output_tokens":82258,"cost_estimate_usd":3.4657121},
    {"contestant":"GPT-5.6 Luna","lane":"A","protocol_correct":16,"exact_cases":1,"invalid_trials":0,"execution_ms":296033,"tool_events":0,"input_tokens":289024,"output_tokens":13697,"cost_estimate_usd":0.371206},
    {"contestant":"Claude Opus 5","lane":"B","protocol_correct":0,"raw_correct":96,"exact_cases":0,"raw_exact_cases":6,"invalid_trials":8,"execution_ms":364977,"tool_events":35,"input_tokens":1489495,"output_tokens":19965,"cost_estimate_usd":4.315915},
    {"contestant":"Claude Haiku 4.5","lane":"A","protocol_correct":0,"exact_cases":0,"invalid_trials":0,"execution_ms":1058605,"tool_events":0,"input_tokens":218476,"output_tokens":98985,"cost_estimate_usd":1.146448},
    {"contestant":"Claude Haiku 4.5","lane":"B","protocol_correct":0,"raw_correct":10,"exact_cases":0,"raw_exact_cases":0,"invalid_trials":8,"execution_ms":2154859,"tool_events":59,"input_tokens":3145023,"output_tokens":145142,"cost_estimate_usd":1.9173406}
  ],
  "artifacts": {
    "protocol": "capability-tournament-v1-protocol.md",
    "results": "capability-tournament-v1-results.md",
    "adversarial_review": "capability-tournament-v1-adversarial-review.md",
    "claude_extension": "capability-tournament-v1-claude-extension-results.md"
  }
}
