EA://INTEL — The Cost of Marking Doubt in Self-Monitoring Systems

An AI system must distinguish between two kinds of wrong information to repair itself. An error is a position you hold confidently but wrongly. An erasure is a position you have flagged as doubtful and therefore do not trust. A code repairs erasures roughly twice or three times faster than it repairs errors because you do not need the first step: finding the damage.

Repairing an error costs two things. You must locate the incorrect position, then fix its value. Repairing an erasure costs only one thing. You already know where the problem is. If your system marks a belief as doubtful, half the repair budget disappears because you skipped the search phase. This mechanical advantage applies to any architecture that tracks confidence intervals or evidence weights.

A page of text with three words missing can usually be restored, because you know which words are gone. The same page with two words silently replaced by wrong data is harder, because you must work out which two were corrupted before fixing them. Marking a belief as doubtful converts an expensive problem into a cheap one. This conversion rate is computable from the structure you choose to represent uncertainty.

The strongest objection states that a system that flags everything has said nothing. That limit is real. The capacity to flag doubt is a bounded budget, not a license for infinite skepticism. A recorded failure teaches where the structure must change or expand. What does your system currently do with a belief it holds at low confidence — mark it, or store it the same as any other?