{
  "schema_version": 2,
  "manifest_ref": "post-training:coba-rl",
  "benchmark": "arithmetic-generate",
  "evaluation_tier": "l1_mechanism_real_teacher",
  "seeds": [
    42,
    43,
    44
  ],
  "algorithm": "coba-rl",
  "teacher": {
    "model_id": "Qwen/Qwen2.5-0.5B-Instruct",
    "model_revision": "7ae557604adf67be50417f59c2c2f167def9a775",
    "actual_calls": 3,
    "cache_hits": 0,
    "request_rate": 1.0,
    "input_tokens": 169,
    "output_tokens": 94,
    "checkpoint_committed": false
  },
  "boundary_cache": {
    "examples_per_seed": 8,
    "samples_per_example": 2,
    "seed_pass_at_k": [
      0.125,
      0.0,
      0.0
    ]
  },
  "capability_boundary_curve": {
    "seed_42": {
      "baseline": {
        "pass@1": 0.0,
        "pass@2": 0.0,
        "pass@4": 0.0
      },
      "final": {
        "pass@1": 0.0,
        "pass@2": 0.0,
        "pass@4": 0.0
      }
    },
    "seed_43": {
      "baseline": {
        "pass@1": 0.041666666666666664,
        "pass@2": 0.041666666666666664,
        "pass@4": 0.0625
      },
      "final": {
        "pass@1": 0.0,
        "pass@2": 0.0,
        "pass@4": 0.020833333333333332
      }
    },
    "seed_44": {
      "baseline": {
        "pass@1": 0.020833333333333332,
        "pass@2": 0.020833333333333332,
        "pass@4": 0.020833333333333332
      },
      "final": {
        "pass@1": 0.0,
        "pass@2": 0.0,
        "pass@4": 0.041666666666666664
      }
    }
  },
  "aggregate_metrics": {
    "baseline_accuracy_mean": 0.0,
    "final_accuracy_mean": 0.0,
    "relative_accuracy": 0.0
  },
  "formal_comparison": true,
  "claim_policy": "one-step three-seed mechanism validation of the real teacher, cache and pass@k paths; no quality-improvement claim",
  "evaluation_protocol": {
    "tier": "l1_mechanism_real_teacher",
    "seeds": [
      42,
      43,
      44
    ],
    "formal_comparison": true,
    "claim_policy": "one-step three-seed mechanism validation of the real teacher, cache and pass@k paths; no quality-improvement claim"
  },
  "provenance": {
    "created_at": "2026-08-20",
    "artifact_path": "docs/post-training/2606.22317-coba-rl/metrics/a30-real-teacher-seeds42-44.json",
    "dataset_fingerprint": "deterministic arithmetic-generate suite",
    "checkpoint_path": "local snapshot (not committed)",
    "resolved_device": "cuda",
    "accelerator": "NVIDIA A30"
  }
}