A Warning Number for AI Checked by Formal Proof: Seeing Drift in the Verifier's Own Results

Abstract

A growing way of making AI trustworthy is to let a formal verifier, such as a Lean proof checker, examine each output or action of the AI and accept or block it. The verifier gives a clear yes or no for each item, but it does not say whether the AI behind it is slowly drifting. This concept note describes a number calculated from the verifier's own stream of results over time. The number rises when the AI has started to recover more slowly from small surprises, so it gives a warning before the failures pile up. Each level of the number has a defined action, decided by the computer inside the platform's own workflow: carry on, check every output of that part, or hold that part for review. The note explains the idea in plain words, how it differs from today's accept-or-block verification and failure counting, what it could mean if it works, and what is not yet known. The limits between the states are to be found on real runs. Patent pending.