ai-audit

Run the four-layer audit: classify logic defects with the ODC taxonomy, verify every dependency against official registries, check for stylistic markers like robotic comments and defensive bloat, and map all prompt constraints against the actual implementation.

Slopsquatting is a supply-chain attack where adversaries register package names that LLMs hallucinate. Detect it by cross-referencing every import against official registries like npm and PyPI, and flagging non-existent versions such as pandas==2.5.0.

AI-generated code is syntactically flawless but often logically hollow, so static analysis passes while algorithmic inefficiencies, missing edge-case handling, and violated prompt constraints remain. The audit targets these semantic and behavioral gaps instead.

Discard the session when code complexity (NLOC/CCN) increases over three iterations without resolving the primary defect. This indicates the model has entered a hallucination loop, and a fresh start is more effective than continued prompting.

The audit is an adversarial manual protocol used when no deterministic evaluator exists for a condition, so it depends on reviewer rigor rather than automated gates. Where a symbolic or test-based check can be built, that deterministic path is preferred.