The Missing Comparator: A Perceptual-Control Loop That Holds Language-Model Claims to an Independent Record — and the Pre-Registered Test That Could Prove It Wrong

Abstract

A language model that answers from an organisation's own documents states supported and unsupported claims in the same confident voice, and nothing in the answer, or usually in the system, tells them apart. An earlier audit traced the same structure across thirty systems: an indicator held in range while the thing it stood for failed, because nothing compared the report with an independent channel. This paper builds that missing comparator into a language-model pipeline and specifies how it will be tested. The system is a runtime control loop around a hosted model, with no retraining. Before an answer is delivered, every factual claim is checked against an independent, read-only record. Two quantities are held to explicit references: - agreement of the answer's claims with the record; - coverage of the question by supported content. A higher level sets the lower references and can never lower the agreement target. Claims the loop cannot support are rewritten, removed, marked unchecked or shown as conflicts; they are never delivered as plain assertions. A hash-chained, per-claim ledger records what was checked, against what, and how. The design controls agreement with records, not truth, and the test is built to show that limit. Evaluation uses the Test for the Controlled Variable from perceptual control theory. The controlled variables are disturbed with their references held fixed. A clamped-feedback test asks whether corrections depend on the loop's own measurements. The loop is compared with verification built from the same components at matched cost. Eight hypotheses are pre-registered, each with the result that would count against it. This version reports no results. Version 2 will report them under this record, whatever they show.