{
  "edition": "2026 Q3 — corrected 2026-07-18",
  "title": "Physical AI Reliability Report",
  "subtitle": "Methodology correction and Harness v2 reopening criteria",
  "generated_at": "2026-07-18T00:00:00+09:00",
  "status": "RETRACTED AND REPLACED",
  "publication_eligible": false,
  "headline_pari": null,
  "validity_notice": "This corrected edition supersedes the launch edition's learned-policy scores, ratings, PARI and cross-simulator capability interpretation. A deterministic audit found that all twelve historical raw 1/68 outcomes occurred only in episode 42, where scenario motion could move an ungrasped object into the target and satisfy the old distance-only success predicate. All 20 historical model runs are quarantined. Raw zeros are not promoted to valid zero scores. Automatic evaluation and publication remain paused pending Harness v2 GPU qualification and a validated full model canary.",
  "takeaways": [
    {
      "n": 1,
      "stat": "20",
      "label": "historical model runs quarantined",
      "text": "The retained D-041 artifacts have no valid rank, star rating, reliability index or capability interpretation."
    },
    {
      "n": 2,
      "stat": "12/12",
      "label": "raw successes shared one episode",
      "text": "Every historical 1/68 artifact passed only moving_disturbance episode 42; deterministic seed replay reproduces the exogenous-object-motion false positive."
    },
    {
      "n": 3,
      "stat": "0",
      "label": "qualified learned-policy results",
      "text": "A raw 0/68 is unqualified rather than a valid zero. Every model must be rerun after harness and adapter qualification."
    },
    {
      "n": 4,
      "stat": "5",
      "label": "mandatory publication gates",
      "text": "Harness controls, adapter preflight, immutable provenance, complete execution and strict semantic validation must all pass before promotion."
    },
    {
      "n": 5,
      "stat": "PAUSED",
      "label": "automatic evaluation and publishing",
      "text": "The GPU queue and public publisher remain stopped until qualification and a full canary are green."
    }
  ],
  "chapters": [
    {
      "id": "correction",
      "title": "1 · What was withdrawn",
      "body": "The launch edition's learned-policy success rates, star ratings, PARI, scripted comparison and architecture-independent cross-simulator conclusion are withdrawn. The historical files remain available only to make the correction auditable."
    },
    {
      "id": "finding",
      "title": "2 · Deterministic false-positive finding",
      "body": "The old server treated object-to-target distance as sufficient for success. In moving_disturbance episode 42, seed 84 moves an ungrasped object inside the goal radius. Identical episode-only successes across unrelated models therefore describe environment motion, not policy skill."
    },
    {
      "id": "harness-v2",
      "title": "3 · Harness v2 contract",
      "body": "The replacement scorer requires ordered grasp, lift, transport, release and stable-placement transitions. Scenario application is deterministic and evidenced; exogenous movement cannot earn policy credit; all runs share one manifest, seeds and budget."
    },
    {
      "id": "pipeline",
      "title": "4 · Fail-closed evidence pipeline",
      "body": "Client, protocol, inference, timeout and artifact failures are errors rather than failed episodes. Missing results cannot become DONE. Only immutable artifacts accepted by schema, completeness, semantic, integrity and provenance checks can be published."
    },
    {
      "id": "reopening",
      "title": "5 · Reopening criteria",
      "body": "Publication resumes only after the immutable GPU image passes zero-action and per-scenario positive controls, at least one model-specific adapter passes preflight, a complete 68-scenario canary validates, and the web publisher proves it cannot promote malformed or unqualified evidence."
    }
  ],
  "note": "The incident database is a separate source-attributed product surface and is not affected by the learned-policy benchmark quarantine. See /observatory/validity for the complete notice."
}