[
  {
    "adapter_name": "aacr-bench",
    "adapter_builders": [
      {
        "name": "Onur Solmaz",
        "email": "2453968+osolmaz@users.noreply.github.com"
      }
    ],
    "original_benchmark": [
      {
        "split": "positive_samples",
        "size": 196,
        "harness": "llm",
        "supported_agents": ["codex", "claude-code", "open-code-review"],
        "adaptable": true,
        "notes": "1,505 reference comments on 196 real pull requests. Matching uses path, diff side, line range, and an LLM semantic-equivalence judge with micro aggregation."
      }
    ],
    "harbor_adapter": [
      {
        "split": "sealed-test",
        "adapted_benchmark_size": 57,
        "parity_benchmark_size": null,
        "parity_sampling_rate": null,
        "registry_benchmark_size": 57,
        "added_agents": ["codex-review", "pi-reviewer"],
        "parity_matching_agents": null,
        "parity_unmatching_agents": null,
        "parity_costs": null,
        "notes": "The immutable offline source corpus contains 57 sealed tasks. Two force-pushed records are excluded because their exact commits and trees are absent from checked public archives. Generated tasks keep references and judge code in the verifier boundary and ship a count-based micro metric."
      }
    ]
  }
]