The green-light fallacy: when agent verification succeeds and the real postcondition does not

The argument

Agentic workflows keep adding automatic checks: tests, lint, type checks, builds, CI, secret scanners, code review, agent review, hashes, manifests and post-task verification. Each one is useful. The fallacy is the step from all of them passing to the system being safe. Unless the global postcondition G was mapped into their coverage, their conjunction says nothing about it — and more green lights can add confidence without adding safety.

The paper separates verification count from verification information gain. Checks that read the same source, share the same abstraction or reuse the same generated test oracle are not independent: n reviewers can amount to one shared blind spot. It distinguishes a detector's false negative from a scope false PASS, where the detector worked correctly but was pointed at the wrong surface, and it insists that NOT MEASURED is a result of its own, never a pass.

Its engineering rule fits on one line: claimed assurance should not exceed demonstrated coverage.