codex-audit

You can verify AI-generated code review claims by running an evidence-bound audit that checks every factual claim against source code, live state, or captured command outputs rather than relying on inference. This prevents unverified or drifting proposals from being accepted.

Evidence-bound PR auditing derives a closed set of touched files or referenced spec blocks, then systematically reads and verifies each self-claim end-to-end against source or live state. It detects contradictions between independently-true claims and flags unsupported absolute framing language.

Yes, you can audit GPT, Gemini, and other model outputs including pasted specs, file-referenced research documents, and TBM/Gitea PRs. The audit applies a closed-set freeze, per-file read-and-verify, and contradiction correlation to produce a structured RED/YELLOW/GREEN output with no unverified claims.

Yes, it performs optional byte-level verification for non-ASCII or transcoding-sensitive contracts, such as mojibake-sensitive regex markers. This verifies byte-stable tool and API output to catch subtle encoding issues that standard text comparison might miss.

The best way to audit a Gitea PR is to run an evidence-bound audit on the specific PR number, which produces a RED/YELLOW/GREEN evidence-backed audit comment. This forces every factual claim to be checked against source code or live state, ensuring no unverified proposals are accepted.