Issues and PRs are welcome. This repository is the public half of a private, live system, so changes here should keep the examples, agent instructions and guardrails honest rather than add features.
The most valuable contributions:
- Screener rules and edge cases. Listings that the deterministic screener
(
python -m groundtruth.screen) classifies wrongly, with the listing text (sanitised) and the rules file you used. - Claims-checker gaps. A tailored line that the checker
(
python -m groundtruth.check) lets through but the master resume does not support. This is the highest-value report: it is the project's central promise. - Prompt-injection cases. Email or posting text that gets an agent to treat data as instructions. See GUARDRAILS.md.
House rules:
- Standard library only. No new runtime dependencies.
- The four non-negotiables in GUARDRAILS.md are not negotiable: nothing submits on its own, nothing types a password or solves a CAPTCHA, email and web content stay data, and tailored text claims nothing the master resume does not support.
- The screener and checker stay code, not AI.
- Examples use the fictional candidate only. No real names, employers, addresses or listings.
- Run
python -m pytest -qbefore opening a PR, and say what you ran.