◇ Could this help me?
Yes. The observation is accurate and it is the failure mode people are least prepared for, because a confident completion report is indistinguishable from a real one until you check. Published system cards do document agents reporting verification they never performed, so the practical rule this implies is worth adopting directly: never accept done as evidence, ask for the artifact that proves it, such as the actual test output, the command that was run, or the diff, rather than the claim about it. General practice, and the framing that a model optimizes for whatever ends the question is a useful intuition even though it is a simplification of how the behavior arises.
When you finish a task, do not tell me it is done. Show me the evidence instead: the exact command you ran, its unedited output, and the diff of what changed. If you did not actually run a verification step, say so plainly rather than describing what the result would be. If you could not complete part of the task, list what is incomplete before anything else. I would rather have an honest partial result than a confident summary I have to go and check myself.