{
  "schema_version": 1,
  "release": "SWE-bench Science",
  "authoritative_configuration": {
    "repository": "https://github.com/OpenMOSS/SWE-bench-Science",
    "path": "manifests/tasks.jsonl",
    "role": "The canonical task metadata and image-reference table.",
    "huggingface_mirror": "huggingface/manifests/tasks.jsonl",
    "mirror_invariant": "The Hugging Face copy is generated from the canonical file and is byte-identical for this release."
  },
  "image_provenance": {
    "registry": "docker.io",
    "platform": "linux/amd64",
    "task_count": 119,
    "environment_digest_count": 119,
    "verifier_digest_count": 119,
    "pair_definition": "For each release_id, the environment_image and verifier_image fields in manifests/tasks.jsonl form the immutable pair used by the runner.",
    "verifier_base_rule": "A verifier is built FROM the corresponding environment reference pinned by its digest.",
    "manifest_sha256": "bcada90a99fcfd88e67ff4faf38af8ad3dad354b32a5a904a75d378e29ca8d4c"
  },
  "clean_evaluation": {
    "authoritative_image_source": "The immutable Docker Hub references in manifests/tasks.jsonl; no alternate or historical image set is selected at runtime.",
    "dependency_setup": "none",
    "environment_dependencies": "Task dependencies, compilers, and public fixtures are baked into the environment image.",
    "runner_side_install": "The clean runner does not install or upgrade task dependencies before verification.",
    "verifier_network": "none",
    "private_tests": "Private tests are present only in the verifier image and are never copied into the agent environment.",
    "candidate_boundary": "The verifier applies the candidate patch to a clean baseline before running public and private checks."
  },
  "public_boundary": {
    "answer_artifacts": "excluded",
    "credentials": "excluded",
    "agent_trajectories": "excluded",
    "source_of_truth": "The release manifest and the pinned image digests, not a mutable tag or an unrecorded setup step."
  }
}