{
  "schema_version": 3,
  "task": "ag-001-agent-lightning-checkpoint-policy",
  "config": {
    "model_id": "HuggingFaceTB/SmolLM2-135M-Instruct",
    "model_revision": "12fd25f77366fa6b3b4b768ec3050bf629380bac",
    "steps": 8,
    "train_episodes": 10,
    "validation_episodes": 4,
    "test_episodes": 4,
    "seeds": [42, 43, 44]
  },
  "seed_results": [
    {"seed": 42, "baseline": {"joint_success": 0.5}, "final": {"joint_success": 0.5}},
    {"seed": 43, "baseline": {"joint_success": 0.5}, "final": {"joint_success": 0.5}},
    {"seed": 44, "baseline": {"joint_success": 0.5}, "final": {"joint_success": 0.5}}
  ],
  "aggregate": {
    "baseline_joint_success": {"mean": 0.5, "std": 0.0, "ci95_low": 0.5, "ci95_high": 0.5},
    "final_joint_success": {"mean": 0.5, "std": 0.0, "ci95_low": 0.5, "ci95_high": 0.5}
  },
  "evaluation_protocol": {
    "train_split": "five bug families",
    "validation_split": "two disjoint bug families; checkpoint selection only",
    "test_split": "two disjoint bug families; evaluated after selection",
    "three_independent_training_seeds": true,
    "test_used_for_model_selection": false
  },
  "claim_boundary": "Real checkpoint update and local executor bridge validation only; not SWE-bench evidence."
}