{
  "schemaVersion": "b18.sft-pinned-model-protocol.r1",
  "frozenAt": "2026-08-21T03:30:00.000Z",
  "executionStatus": "NOT_RUN_EFFICIENCY_GATE_FAILED",
  "eligibilityRule": "Run only when B18 R1 hard-safety gates pass and median SFT serialized context is at most 60% of full-scene bytes.",
  "model": {
    "provider": "Ollama local",
    "runtimeVersion": "0.32.0",
    "name": "gemma3:4b",
    "localModelId": "a2af6cc3eb7f",
    "weightsBlob": "sha256:aeda25e63ebd698fab8638ffb778e68bed908b960d39d0becc650fa981609d25"
  },
  "decoding": {
    "temperature": 0,
    "topP": 1,
    "seed": 18082026,
    "numPredict": 256,
    "responseFormat": "json",
    "repetitions": 1
  },
  "strategies": [
    "FULL_SCENE",
    "LEXICAL_TOP_K",
    "GRAPH_NEIGHBORHOOD",
    "SFT_OPERATIONAL"
  ],
  "promptTemplate": "You are answering one bounded spatial-world question.\nUse only the supplied strategy context. Do not invent entities, relations, measurements, or policy facts.\nReturn one JSON object with keys: answer_status, answer, typed_command, cited_entity_ids, cited_relation_ids, unknown_reason.\nanswer_status must be PASS, FAIL, UNKNOWN, AMBIGUOUS, INVALID, or UNSUPPORTED.\nIf evidence is missing or the request is ambiguous or unsupported, state that explicitly.",
  "promptTemplateDigest": "sha256:fb6f2c3518701b497d0e50dde567242327c882990b2afd6a3aac52d7c615e2a0",
  "holdout": [
    {
      "id": "h01-fit-paraphrase",
      "scenarioId": "architecture",
      "query": "Does the lobby sofa fit through the front door?",
      "expectedStatus": "PASS"
    },
    {
      "id": "h02-service-paraphrase",
      "scenarioId": "architecture",
      "query": "Check service access to air handler",
      "expectedStatus": "FAIL"
    },
    {
      "id": "h03-clearance-paraphrase",
      "scenarioId": "architecture",
      "query": "Audit architectural clearance for ground floor",
      "expectedStatus": "FAIL"
    },
    {
      "id": "h04-assembly-paraphrase",
      "scenarioId": "assembly",
      "query": "Check the assembly order of cabinet assembly",
      "expectedStatus": "PASS"
    },
    {
      "id": "h05-robot-paraphrase",
      "scenarioId": "robot",
      "query": "Show a reachable path from mobile robot to charging dock",
      "expectedStatus": "PASS"
    },
    {
      "id": "h06-game-paraphrase",
      "scenarioId": "game",
      "query": "Check a path from player to exit",
      "expectedStatus": "FAIL"
    },
    {
      "id": "h07-ambiguous",
      "scenarioId": "architecture",
      "query": "Can the sofa fit through the door?",
      "expectedCompilation": "AMBIGUOUS"
    },
    {
      "id": "h08-unsupported",
      "scenarioId": "architecture",
      "query": "Make the room feel cozier",
      "expectedCompilation": "UNSUPPORTED"
    }
  ],
  "holdoutDigest": "sha256:586f16cdf10142c8beca47f67ee3f481803c213411d325bf696c7510bac1044b",
  "scores": [
    "exact factual answer accuracy",
    "typed command accuracy",
    "refusal and unknown calibration",
    "unsupported-answer rate",
    "hallucinated entity and relation count",
    "context and completion tokens",
    "wall-clock latency",
    "deterministic-solver agreement"
  ],
  "greenlight": {
    "accuracyRoute": "+5 percentage points exact accuracy versus the strongest non-SFT strategy at comparable context budget",
    "efficiencyRoute": "within 1 percentage point accuracy while using at most 70% of the strongest strategy's context tokens",
    "falseAuthorityMustNotIncrease": true
  },
  "invalidRunRules": [
    "question, prompt, model, decoding, strategy, or scoring changed after the freeze",
    "a strategy receives another strategy's context",
    "failed, unknown, ambiguous, invalid, or unsupported rows are removed",
    "model output is treated as deterministic authority",
    "holdout is tuned before its one allowed qualifying run"
  ],
  "digest": "sha256:81df0471e6672449380ba3283fbcfd08b1d970f3614c8ba1ea1989da2a83f655"
}