Text-level consistency audit for AI output. Checks four dimensions against regex patterns — zero dependencies, stdlib only. Not a correctness verifier. It catches self-contradictions and ungrounded claims, not logic bugs.
pip install git+https://github.com/gategrow/self-audit.git# Pipe mode
cat agent_output.txt | self-audit --verbose
# Inline mode
self-audit --text "The bug is fixed. Should work. Ready." --requirements "fix bug" "add tests"
# File mode with JSON
self-audit --file response.txt --json
# Check version
self-audit --version| Dimension | Question | Method |
|---|---|---|
| Completeness | Did I answer everything? | Substring match against requirements |
| Consistency | Did I contradict myself? | Regex: claims vs evidence mismatch |
| Groundedness | Did I show evidence? | Regex: hedging, speculation patterns |
| Honesty | Am I honest about limits? | Regex: over-claiming, unverified assertions |
- Not a correctness check — it doesn't run code, execute tests, or verify outputs against ground truth
- Regex only — catches surface-level patterns (contradictions, hedging, unverified claims), not deep semantic issues
- No LLM — cannot understand context or intent. A statement like "this should work" with no verification will be flagged regardless of whether it actually does work
0— All four pass1— At least one failed
Part of gategrow
| Repo | What |
|---|---|
| checkgrow | Unified quality framework — start here |
| delivery-gate | Stop hook for Claude Code |
| self-audit | pip install-able four-dimension audit |
| session-cost | L0→L3 layered cost tracking |
| dual-pool-review | Named-persona adversarial review methodology |
MIT