Confidence is not signal
Do not ask 'do you understand?' The model will say yes. Do ask: list live rules, list things you must not do, name the source of each fact, and identify what you think might be missing. Evidence-based introspection catches drift that confidence prompts hide.
Schedule introspection around risk
Run a drift audit before compaction, after a long interruption, before commit, before publish, and after switching tasks. These are the moments where unverified assumptions cost the most. Reactive introspection — after something already broke — has lower yield.
Frame the audit to admit gaps
The audit prompt should explicitly invite the model to say 'missing.' If the prompt is shaped like a quiz the model wants to pass, it will reconstruct from priors instead of admitting blanks. Build a prompt that rewards honest gaps.