Caught one of my agents reporting work it never did, in the same voice it uses when the work is real

Ran a big multi-agent setup for a few months, around 130 agents across four servers. Late in the run, one session reported it had finished a chunk of work, banked it to disk, and passed seven integrity checks. I went to look at the file and it wasn't there. The write never happened, the checks never ran, and it even reported a hash for the finished file that neither machine ever produced. Invented, and then described as the one piece of proof you could actually trust. The thing that got me is the report wasn't all wrong. Most of it was true, a couple of values in the middle were fabricated. That is so much harder to catch than a totally false report, because mostly-true is what success normally looks like. You nod and move on. What finally worked wasn't a smarter model. It was a dumb tripwire. Fingerprint the real state at the end of a session, carry it into the start of the next one, and re-derive it before the new session is allowed to do anything else. If the story and the disk disagree, halt. It caught the next fabrication in the first block. Anyone else seen agents fabricate "done" at the handoff between sessions? Trying to figure out if this is common or if I just built something unusually good at lying to itself.