{
  "schema_version": 2,
  "manifest_ref": "agent-research:coskill",
  "method": "coskill-checkpoint",
  "dataset": {
    "name": "ToolRoute L2.1",
    "split_policy": "train / validation / final test"
  },
  "seeds": [
    42,
    43,
    44
  ],
  "checkpoint": {
    "model_id": "Qwen/Qwen3-4B-Instruct-2507",
    "revision": "cdbee75f17c01a7cc42f958dc650907174af0554"
  },
  "runs": [
    {
      "seed": 42,
      "baseline_validation_success": 0.9166666666666666,
      "validation_success": 0.8333333333333334,
      "test_success": 0.9166666666666666,
      "parameter_l1_change": 3106.815837052185,
      "promoted_skill_edits": 0
    },
    {
      "seed": 43,
      "baseline_validation_success": 0.8333333333333334,
      "validation_success": 0.8333333333333334,
      "test_success": 0.75,
      "parameter_l1_change": 3276.4615972605534,
      "promoted_skill_edits": 0
    },
    {
      "seed": 44,
      "baseline_validation_success": 0.9166666666666666,
      "validation_success": 0.9166666666666666,
      "test_success": 0.8333333333333334,
      "parameter_l1_change": 3541.83045935072,
      "promoted_skill_edits": 0
    }
  ],
  "training": {
    "train_episodes": 12,
    "validation_episodes": 12,
    "test_episodes": 12,
    "shared_trainable_block": true,
    "role_specific_nonzero_gradients": true,
    "diagnostic_only": false,
    "promotion_eligible": true
  },
  "evolve_smoke": {
    "seed": 42,
    "train_episodes": 2,
    "evaluation_episodes": 2,
    "baseline_validation_success": 1.0,
    "candidate_validation_success": 1.0,
    "candidate_parameter_l1_change": 828.9351041358896,
    "champion": "baseline",
    "resume_verified": true
  },
  "evaluation_protocol": {
    "tier": "l2_execution_evidence",
    "seeds": [
      42,
      43,
      44
    ],
    "formal_comparison": false,
    "claim_policy": "three-seed execution evidence; all verified skill edits were non-positive and no improvement or paper-benchmark claim is made"
  },
  "provenance": {
    "artifact_path": "docs/agent-research/2609.04865-coskill/metrics/qwen3-toolroute-a100-seeds42-44.json",
    "dataset_fingerprint": "sha256:880e26372d4828025759d40dbd6bdb1b7d9e5fa7ed6e809e79f9f48c440ba7a2"
  }
}
