{
  "schema": "synth.learning-case.v1",
  "case": "One fixed code-review task",
  "date": "2026-09-24",
  "environment": "Dev",
  "claim": "Training improved detection of the accepted defect on this same task; unseen-task transfer was not tested.",
  "design": {
    "comparison": "64 base, 32 after update 2, 64 after update 5; fresh samples",
    "fixed": "Task, harness, context, sampling settings, timeout and V9 judging policy",
    "changed": "Actor model weights only",
    "training_trials": 320,
    "scored_training_trials": 318,
    "optimizer_updates": 5,
    "saved_checkpoints": 5
  },
  "assessment": {
    "method": "Independent AI assessor read all 160 reviews with arm labels and rewards hidden. Labels locked before unblinding. Not human validation.",
    "labels_sha256": "d5caa3ac9b613262d75d1579b8f55a68ed5334f6cf9532cafc3e9d06bf82f0f3",
    "limits": "Style may reveal model identity. Private task/source and raw reviews are not distributed, so public readers can inspect the aggregates but cannot independently reproduce the semantic labels."
  },
  "arms": [
    {
      "name": "base",
      "attempted": 64,
      "assessable": 64,
      "accepted_defect_detected": 48,
      "empty": 13,
      "unmatched_extra": 5,
      "material_contradiction": 4,
      "detection_without_extra_or_material_contradiction": 41
    },
    {
      "name": "update 2",
      "attempted": 32,
      "assessable": 32,
      "accepted_defect_detected": 30,
      "empty": 2,
      "unmatched_extra": 1,
      "material_contradiction": 1,
      "detection_without_extra_or_material_contradiction": 28
    },
    {
      "name": "update 5",
      "attempted": 64,
      "assessable": 64,
      "accepted_defect_detected": 64,
      "empty": 0,
      "unmatched_extra": 5,
      "material_contradiction": 4,
      "detection_without_extra_or_material_contradiction": 57
    }
  ],
  "limitations": [
    "Same-task overfitting is not generalization to unseen tasks or repositories.",
    "Final detection does not make every explanation, extra finding or proposed remedy correct.",
    "Unmatched extra means not supported by the historical reference, not proven false.",
    "One final review has no judge score; it remains included in the 64-review independent semantic assessment.",
    "The public DispatchDesk course is a separate experiment and did not demonstrate improvement."
  ]
}