# AACR-Bench Harbor

- Keep this repository independent. Do not turn it into a fork, submodule checkout, or vendored copy of AACR-Bench or Harbor.
- Follow the current Harbor adapter, task, metric, and publishing documentation before changing the format.
- Treat adapter source as canonical. Never hand-edit generated task directories.
- Preserve AACR-Bench's review target, matching stages, raw counts, and micro-aggregated metrics unless a documented change intentionally replaces them.
- Keep reference findings and judge code outside the agent-visible environment.
- Pin upstream data, repositories, dependencies, agents, models, prompts, and judge configuration.
- Fail closed on judge errors, malformed findings, missing outputs, and unsupported records. Never report mock-judge results as benchmark results.
- Use strict Python typing. Avoid `Any`; validate unknown external input at the boundary.
- Keep domain behavior separate from files, networks, subprocesses, clocks, and process state.
- Add tests for every behavior change. Do not add dependencies when the standard library suffices.
- Keep mutation tooling available, but run mutation tests only when explicitly requested or investigating test strength.
- Before finishing, run `scripts/check.sh`, `uv run slophammer-py check . --execute`, `npx -y @simpledoc/simpledoc check`, and `git diff --check`.
- Do not launch paid parity or benchmark runs before the required canary, cost estimate, durable resume proof, and approval.
- Never commit credentials, local authentication state, generated runs, cloned repositories, or provider secrets.
- Slophammer guidance: https://github.com/osolmaz/slophammer/blob/main/docs/AGENT_ENTRYPOINT.md