{
  "baseline_sha256": "984ff1aaed71652d9f53c91074e0787b3386bc3ae41900b4b830e6211805fc64",
  "candidate_sha256": "1cdbff7ac98dfbc4abeb00e2f0a7b2eea124674d4eebc9fe5bee0357ea8b4502",
  "estimated_gpu_usd": 1.1973,
  "families": {
    "boolean": {
      "baseline": {
        "n": 40,
        "outcomes": 8,
        "plans": 8
      },
      "candidate": {
        "n": 40,
        "outcomes": 40,
        "plans": 40
      }
    },
    "credentials": {
      "baseline": {
        "n": 40,
        "outcomes": 0,
        "plans": 0
      },
      "candidate": {
        "n": 40,
        "outcomes": 40,
        "plans": 40
      }
    },
    "dates": {
      "baseline": {
        "n": 40,
        "outcomes": 25,
        "plans": 16
      },
      "candidate": {
        "n": 40,
        "outcomes": 40,
        "plans": 40
      }
    },
    "dotenv": {
      "baseline": {
        "n": 40,
        "outcomes": 0,
        "plans": 0
      },
      "candidate": {
        "n": 40,
        "outcomes": 40,
        "plans": 40
      }
    },
    "formats": {
      "baseline": {
        "n": 40,
        "outcomes": 4,
        "plans": 4
      },
      "candidate": {
        "n": 40,
        "outcomes": 40,
        "plans": 40
      }
    },
    "refusal": {
      "baseline": {
        "n": 40,
        "outcomes": 12,
        "plans": 12
      },
      "candidate": {
        "n": 40,
        "outcomes": 40,
        "plans": 40
      }
    }
  },
  "gpu_deleted": true,
  "limitations": [
    "synthetic same-teacher language review",
    "135 filename validation rows are literal-plan diagnostics only",
    "no app latency or independent human evaluation"
  ],
  "numeric_gate": false,
  "older_outcome_improvements": 6,
  "older_outcome_regressions": 12,
  "optimizer_steps": 1213,
  "promoted": false,
  "regressed_cases": 15,
  "reports": {
    "baseline": {
      "filename-validation": {
        "literal_exact": 125,
        "n": 135,
        "outcomes": null,
        "plans": null
      },
      "retention": {
        "literal_exact": null,
        "n": 160,
        "outcomes": 145,
        "plans": 129
      },
      "transfer": {
        "literal_exact": null,
        "n": 240,
        "outcomes": 49,
        "plans": 40
      }
    },
    "candidate": {
      "filename-validation": {
        "literal_exact": 126,
        "n": 135,
        "outcomes": null,
        "plans": null
      },
      "retention": {
        "literal_exact": null,
        "n": 160,
        "outcomes": 139,
        "plans": 125
      },
      "transfer": {
        "literal_exact": null,
        "n": 240,
        "outcomes": 240,
        "plans": 240
      }
    }
  },
  "schema": "shlex-live-repair-public-v1",
  "training_rows": 19408,
  "validation_rows": 2046
}
