The Admissibility Review: A No-Over-Claim Soundness Guarantee for Language-Model Assertions

Abstract

Decision support systems, and the AI components inside them, are built to produce conclusions; whether a conclusion may be asserted at the strength it is stated - given the evidence behind it - is governed, if at all, after the fact and detachably. In high-stakes decisions on physical assets under sparse evidence, the result is over-claim: conclusions that outrun their evidence enter the decision record as if verified, where they are neither reproducible nor auditable. We invert the pipeline, making the evidence state sovereign over the assertion. We fix a computable ceiling on the force a conclusion may take as a function of its evidence maturity, and enforce that ceiling at emission - a check we frame as an admissibility review of the evidence behind a decision. We formalize this as a monotone morphism from an evidence-maturity lattice to an assertion-force lattice, prove a no-over-claim soundness theorem - no emitted claim exceeds the meet of its evidence dependencies - and identify it as the integrity dual of information-flow non-interference. We describe a reference implementation that enforces the ceiling default-on for physical-asset capital decisions, confining the stochastic generator to a bounded component so the guarantee is architectural, and evaluate it by demonstration, ablation, and outcome-vindication on real decisions. The guarantee is soundness, not completeness: the system may under-claim, never over-claim. It does not certify that the evidence is correct - it makes the evidentiary chain explicit, so over-claims are localizable and challengeable rather than hidden in the decision record.