Interpretability for Oversight

As models get more capable, a gap opens: the outputs become harder to evaluate than the computation that produced them. A model can produce a correct-looking proof with a subtle flaw, eloquent advice with hidden agenda, or a backdoored solution dressed as a clean one. Output-only oversight — judging the answer — fails exactly where capability exceeds the evaluator.

Interpretability is the bet that the internals carry the evidence: the deceptive plan exists as activations before the deceptive output exists as tokens. If you can read the feature for 'planning to mislead' or 'knows this is false,' you don't need to out-argue the model — you need to out-read it.