AI-generated code can pass all tests while still implementing the wrong feature — because tests verify behavior, not intent. When an LLM agent fills gaps in a prompt with plausible guesses and then writes tests for its own output, the result is green CI with wrong functionality. The solution is a separate verification layer that checks implementation against a human-approved spec with explicit acceptance criteria. Aviator's experiment showed a verification agent checking 65 acceptance criteria in 6 minutes, catching 4 intent mismatches that no test would have flagged. The proposed workflow separates concerns: humans own spec approval, tests own behavior, and automated verification owns agreement between implementation and spec on every revision.
