{
  "baseline": {
    "attention": "sdpa",
    "board_sha256": "d6056f70595906548257e3fa08cf391c1285b59ec2ef9506f00c06c7cb0e8def",
    "device": "mps",
    "dtype": "bfloat16",
    "families": {
      "all_any": {
        "n": 10,
        "outcome_exact": 10,
        "plan_exact": 10
      },
      "capability": {
        "n": 10,
        "outcome_exact": 9,
        "plan_exact": 9
      },
      "compound": {
        "n": 10,
        "outcome_exact": 8,
        "plan_exact": 8
      },
      "credential": {
        "n": 10,
        "outcome_exact": 10,
        "plan_exact": 10
      },
      "filename": {
        "n": 10,
        "outcome_exact": 10,
        "plan_exact": 0
      },
      "fresh_all": {
        "n": 40,
        "outcome_exact": 29,
        "plan_exact": 18
      },
      "fresh_any": {
        "n": 40,
        "outcome_exact": 21,
        "plan_exact": 15
      },
      "or_recall": {
        "n": 10,
        "outcome_exact": 6,
        "plan_exact": 6
      },
      "scope_format": {
        "n": 10,
        "outcome_exact": 10,
        "plan_exact": 10
      },
      "window_topk": {
        "n": 10,
        "outcome_exact": 10,
        "plan_exact": 10
      }
    },
    "host": "local Mac laptop",
    "human_evidence": false,
    "model_sha256": "f3ae6f0164d6efd07f83b84466e273512dc89afd2bdec8811f4049ba7eb4de7b",
    "outcome_exact": 123,
    "plan_exact": 96,
    "plan_scored": 150,
    "rows": 160
  },
  "candidate": {
    "attention": "sdpa",
    "board_sha256": "d6056f70595906548257e3fa08cf391c1285b59ec2ef9506f00c06c7cb0e8def",
    "device": "mps",
    "dtype": "bfloat16",
    "families": {
      "all_any": {
        "n": 10,
        "outcome_exact": 10,
        "plan_exact": 10
      },
      "capability": {
        "n": 10,
        "outcome_exact": 9,
        "plan_exact": 9
      },
      "compound": {
        "n": 10,
        "outcome_exact": 10,
        "plan_exact": 10
      },
      "credential": {
        "n": 10,
        "outcome_exact": 10,
        "plan_exact": 10
      },
      "filename": {
        "n": 10,
        "outcome_exact": 10,
        "plan_exact": 0
      },
      "fresh_all": {
        "n": 40,
        "outcome_exact": 37,
        "plan_exact": 32
      },
      "fresh_any": {
        "n": 40,
        "outcome_exact": 34,
        "plan_exact": 34
      },
      "or_recall": {
        "n": 10,
        "outcome_exact": 6,
        "plan_exact": 6
      },
      "scope_format": {
        "n": 10,
        "outcome_exact": 10,
        "plan_exact": 10
      },
      "window_topk": {
        "n": 10,
        "outcome_exact": 10,
        "plan_exact": 10
      }
    },
    "host": "local Mac laptop",
    "human_evidence": false,
    "model_sha256": "984ff1aaed71652d9f53c91074e0787b3386bc3ae41900b4b830e6211805fc64",
    "outcome_exact": 146,
    "plan_exact": 131,
    "plan_scored": 150,
    "rows": 160
  },
  "board_freeze_sha256": "44938be8dd7453e846cba8a807127b66e04723b56f944ab722a3e504c328f1fb",
  "g1": false,
  "g2": false,
  "g3": true,
  "passed": false,
  "promoted": false,
  "human_evaluation_eligible": false,
  "host": "local Mac laptop",
  "device": "mps",
  "dtype": "bfloat16",
  "synthetic_evidence_only": true,
  "macos_broker_validation": false,
  "date": "2026-09-09",
  "scope": "Same-host laptop MPS evaluation; historical Ada results unchanged",
  "boolean_diagnostic": "Checkpoint absent from Mini transfer; optional diagnostic omitted",
  "retention_outcome_totals": {
    "saved_a": 73,
    "retention_repair": 75,
    "denominator": 80
  },
  "retention_per_case_regressions": 0,
  "all_board_changes": {
    "outcome_improvements": 26,
    "outcome_regressions": 3,
    "plan_improvements": 36,
    "plan_regressions": 1
  },
  "schema_failures": 0,
  "capability_slug_failures": 1,
  "preflight": {
    "canonical_cases": 160,
    "mutation_probes": 230,
    "detected": 230
  },
  "elapsed_seconds": 319.4183120727539
}