źlab Research
The formal science behind the framework.
Our decision reviews are not a heuristic. They instantiate a formal method with a proven guarantee: a system can be built that never asserts a conclusion more strongly than its evidence admits.
The failure mode
In high-stakes decisions on physical assets, conclusions routinely outrun their evidence — a benchmark estimate stated as a verified saving. It enters the record as if proven, and no one can tell which claim is load-bearing.
The contribution
We define a computable ceiling that maps the maturity of the evidence to the strongest claim it admits, and enforce it at the moment a conclusion is stated. We call the resulting check an admissibility review.
The guarantee
We prove a no-over-claim soundness theorem: no emitted claim exceeds the weakest link of its evidence. It is a property of the construction, not a tendency — it holds in every case, by design.
The evidence
On real physical-asset decisions, removing the enforcement lets over-claims through; with it engaged, none survive (a paired ablation, p < 10⁻⁵). The refusals also track decisions reality later refuted.
We are precise about scope: the guarantee is that the evidentiary chain is explicit and bounded, so over-claims can be located and challenged — not that the evidence is correct. And it governs what may be asserted, not which action you should take.
Read it yourself
The Admissibility Review: A No-Over-Claim Soundness Guarantee for Language-Model Assertions
David Lagarejo · DOI 10.5281/zenodo.21955291
Preprint · not peer reviewed · published 15 August 2026
For researchers, auditors, and technical buyers