{
  "benchmarks": [
    {
      "id": "artifact_entrypoint_parity",
      "name": "Artifact verifier canonical vs wrapper parity",
      "command": "bash benchmarks/entrypoint_contract_check.sh",
      "workdir": ".",
      "timeout_s": 120
    },
    {
      "id": "artifact_verifier_determinism",
      "name": "Deterministic verifier parity matrix",
      "command": "python3 benchmarks/verify_benchmark_parity.py",
      "workdir": ".",
      "timeout_s": 180
    },
    {
      "id": "coverage_gap_audit",
      "name": "Coverage gap audit refresh",
      "command": "python3 benchmarks/coverage_gap_audit.py > benchmarks/audit/coverage_gap_audit_evidence.txt",
      "workdir": ".",
      "timeout_s": 120
    },
    {
      "id": "class_trend_generation",
      "name": "Class trend generation",
      "command": "python3 benchmarks/generate_class_trends.py",
      "workdir": ".",
      "timeout_s": 120
    }
  ]
}