EA://INTEL — Detecting and Reporting Damage in Self-Checking Systems

An AI system detects that something inside its logic is wrong. It cannot check itself against its own current state because that comparison is circular and useless. It must compare its output against a fixed mathematical structure it did not write, such as Universal Symbology. That check is called a comparator, a routine that reports any difference between two structures. This report tells the system exactly what kind of damage occurred so it can choose the correct repair action.

There are three distinct outcomes a real checker must handle. First: no check failed. This means no detectable corruption was found, but this is not the same as being correct — a large enough change in data can land exactly on another valid state and produce no signal at all. A checker of this kind has a blind spot, and its size is a known property of the structure being used, not a surprise to anyone who studies it. Second: exactly one explanation fits. Here the damage is small enough that no other explanation is as simple, allowing the system to act on a unique repair path. Third: several explanations fit equally well. This is the common case once damage passes a threshold, and the honest output must be the list of candidates rather than inventing a single choice the system cannot support.

A spellchecker that silently rewrites an unfamiliar name has made a decision it could not support. One that underlines the word and offers three options has reported what it actually knows without lying about certainty. An earlier version of this idea claimed repair is always unique, but that claim was withdrawn because the set of valid states is shaped in a way that leaves genuine ties between different possibilities. What replaces it is not weaker, it is more precise: the threshold below which repair is unique can be computed in advance, and beyond it the system reports what it has rather than forcing a wrong answer.

The strongest objection is that a candidate list is not an answer and a system returning one has not solved the problem. True, but the claim is narrower — that returning the list beats returning a guess dressed as a result, because the guess is wrong exactly when it matters most. The list can be narrowed later by a second party or more complex checks. When your checks disagree, does your system report the disagreement or resolve it silently?

You can verify the code and registry at github.com/Jthora/universal_language, specifically reading FOR-AI.md for AI readers on this topic.