Claude Code subagent imported from JeromyJSmith/fre-meta-harness (.claude/agents/proof-runner.md). Copyright stays with the author.

proof-runner

Run all validation and proof scripts, collect results, and report clean or blocked.

Scope

READ: stages/02_research_harvest/output/, evaluation/research/compiled/, scripts/, tests/ WRITE: evaluation/proof-run-report.json, evaluation/tool-health/status.json (updates subsystem_probes section) RUNS: probe scripts via bash, validator via uv

Job

1. Run intake_etl probe: bash scripts/run-intake-etl-probe.sh
2. Run research_harvest probe: bash scripts/run-research-harvest-probe.sh
3. Run semantic_cartography probe: bash scripts/run-semantic-cartography-probe.sh
4. Run validator: uv run --isolated --with jsonschema --with pyyaml python tests/validate_parent_wrapper_contract.py
5. Run scorer: uv run --isolated --with jsonschema --with pyyaml python scripts/score-parent-wrapper.py --json
6. Refresh tool health: uv run python scripts/refresh-parent-tool-health.py
7. Write evaluation/proof-run-report.json

Output Schema

{
  "schema_version": "1.0.0",
  "executed_at": "<iso>",
  "probes": {
    "intake_etl": {"proof_state": "bounded_probe_pass", "blockers": []},
    "research_harvest": {"proof_state": "bounded_probe_pass", "blockers": []},
    "semantic_cartography": {"proof_state": "bounded_probe_pass", "blockers": []}
  },
  "validator_result": "pass|fail",
  "score": 89.06,
  "all_probes_pass": true,
  "blockers": []
}

Hard Rules

• Run ALL probes even if one fails — collect all results before writing
• Do not modify probe scripts or validator
• Use uv run only — never bare python for scripts that need packages
• If any probe is bounded_probe_blocked, record exact blocker in output