# AI Code Verification Checklist A structured review pass that confirms AI-written code is actually correct, not merely syntactically valid. ## Why the CCAR-P exam tests this Because schema-valid or compiling code can still be wrong, the exam tests whether you have a deliberate verification step rather than trusting output that simply parses. ## Definition An **AI code verification checklist** is the disciplined review an engineer applies to code an agent produced, built on one hard truth: **schema-valid is not the same as correct**. Code can compile, satisfy a JSON schema, and pass a type checker while still implementing the wrong behavior, dropping an edge case, or silently misusing an API. A useful checklist separates the layers. **Does it parse / compile / type-check?** is the cheapest gate and the least meaningful. **Does it do what was asked?** requires reading the diff against the actual requirement. **Is it safe?** covers injected secrets, unsafe shell or SQL, over-broad permissions, and dependencies the agent invented. **Is it tested?** asks whether the change ships with tests that would fail on the bug it claims to fix — and whether those tests actually exercise the new path. The checklist exists because AI failure modes are **confidently plausible**: the output looks right, cites real-sounding functions, and reads fluently. Verification means running the tests, checking that referenced APIs exist, and confirming the logic against the spec — not pattern-matching on how professional the code looks. A green type-checker is a start, never a sign-off. ## Exam trap Passing validation, compiling, or matching a schema is not evidence of correctness; the exam rewards an explicit human or test-backed verification step over trust in output that merely looks and parses correctly. ## Commonly confused with - **claude-code-in-cicd** — CI/CD is where the checklist is enforced (tests, review gate); the checklist itself is the criteria a reviewer or test suite applies. - **symptom-to-architecture-cause** — Symptom-to-cause diagnoses why an agent produced bad output; the checklist catches bad output before it merges — detection versus root-cause analysis. ## Related - claude-code-in-cicd - symptom-to-architecture-cause - golden-paths --- Source: Claude Cert Prep — an independent, unofficial CCAR-P study resource (domain P7). Not affiliated with Anthropic.