{
  "generated_at": "2026-03-04T13:13:50Z",
  "root_dir": "/mnt/c/_organized_non_vital/ai_research",
  "manifest": "/mnt/c/_organized_non_vital/ai_research/benchmarks/benchmark_manifest_core.json",
  "run_id": "2026-03-04",
  "summary": {
    "total": 4,
    "passed": 4,
    "failed": 0
  },
  "benchmarks": [
    {
      "id": "artifact_entrypoint_parity",
      "name": "Artifact verifier canonical vs wrapper parity",
      "status": "passed",
      "pass": 1,
      "runtime_sec": 0.3828,
      "metric": {
        "return_code": 0,
        "stdout_bytes": 455,
        "stderr_bytes": 0
      },
      "command": "bash benchmarks/entrypoint_contract_check.sh",
      "workdir": "/mnt/c/_organized_non_vital/ai_research",
      "return_code": 0,
      "stdout_log": "/mnt/c/_organized_non_vital/ai_research/benchmarks/runs/2026-03-04/artifact_entrypoint_parity.stdout.log",
      "stderr_log": "/mnt/c/_organized_non_vital/ai_research/benchmarks/runs/2026-03-04/artifact_entrypoint_parity.stderr.log"
    },
    {
      "id": "artifact_verifier_determinism",
      "name": "Deterministic verifier parity matrix",
      "status": "passed",
      "pass": 1,
      "runtime_sec": 0.5377,
      "metric": {
        "return_code": 0,
        "stdout_bytes": 0,
        "stderr_bytes": 0
      },
      "command": "python3 benchmarks/verify_benchmark_parity.py",
      "workdir": "/mnt/c/_organized_non_vital/ai_research",
      "return_code": 0,
      "stdout_log": "/mnt/c/_organized_non_vital/ai_research/benchmarks/runs/2026-03-04/artifact_verifier_determinism.stdout.log",
      "stderr_log": "/mnt/c/_organized_non_vital/ai_research/benchmarks/runs/2026-03-04/artifact_verifier_determinism.stderr.log"
    },
    {
      "id": "coverage_gap_audit",
      "name": "Coverage gap audit refresh",
      "status": "passed",
      "pass": 1,
      "runtime_sec": 0.0973,
      "metric": {
        "return_code": 0,
        "stdout_bytes": 0,
        "stderr_bytes": 0
      },
      "command": "python3 benchmarks/coverage_gap_audit.py > benchmarks/audit/coverage_gap_audit_evidence.txt",
      "workdir": "/mnt/c/_organized_non_vital/ai_research",
      "return_code": 0,
      "stdout_log": "/mnt/c/_organized_non_vital/ai_research/benchmarks/runs/2026-03-04/coverage_gap_audit.stdout.log",
      "stderr_log": "/mnt/c/_organized_non_vital/ai_research/benchmarks/runs/2026-03-04/coverage_gap_audit.stderr.log"
    },
    {
      "id": "class_trend_generation",
      "name": "Class trend generation",
      "status": "passed",
      "pass": 1,
      "runtime_sec": 0.179,
      "metric": {
        "return_code": 0,
        "stdout_bytes": 0,
        "stderr_bytes": 339
      },
      "command": "python3 benchmarks/generate_class_trends.py",
      "workdir": "/mnt/c/_organized_non_vital/ai_research",
      "return_code": 0,
      "stdout_log": "/mnt/c/_organized_non_vital/ai_research/benchmarks/runs/2026-03-04/class_trend_generation.stdout.log",
      "stderr_log": "/mnt/c/_organized_non_vital/ai_research/benchmarks/runs/2026-03-04/class_trend_generation.stderr.log"
    }
  ]
}