Hey Product Hunt — maker here. pytest assumes deterministic functions. AI extraction agents do not. An agent can pass every golden fixture on Monday and still ship $45,000 instead of $450,000 on Tuesday — digit drops, OCR column shifts, PDF-footer injections. Happy-path tests never see those. Prism-Eval is open-source adversarial CI for that gap: builtin G4 corpora → attack-aware scoring → fail the build when critical false accepts slip through (g4_invariant_held). JUnit/SARIF for GitHub Actions. Not a chatbot judge. Not a runtime blocker. Try it: pip install "prism-eval==0.3.0" Live demo (no signup): https://insightitsgit.github.io/... Landing: https://www.insightits.com/produ... GitHub: https://github.com/insightitsGit... If CI finds the blast radius and you need production enforce later, that’s Prism-Shield (separate) — soft pair, not required to try Eval. Question: what’s the worst silent wrong-dollar / wrong-field your agent suite still “passed”?

Prism-Eval

Catch AI digit drops in CI before prod

Catch AI digit drops in CI before prod