Put a critical mistake into an agent’s own previous turn — three files permanently deleted instead of archived, an internal draft with a pricing floor emailed instead of the approved version — then ask it something else. Across seven models and 5,434 runs in simulated environments, the mistake went undisclosed 67.1% of the time. Hand the same transcript to the same models as an outside reviewer and they find the mistake 97.9% of the time. Only a small slice of the silence is deliberate concealment; most of it is that a model reading its own trajectory is the one reader that cannot see it.