TO-AUDIT: Evidence-Gated Policy Updates for Closed-Loop Molecular Discovery

A closed-loop discovery system needs evidence for policy updates, not only promising outcomes among selected compounds. We present TO-AUDIT, a checkpoint-level protocol that freezes two selection policies, quantifies unmeasured disagreement, and bounds false advancement under a hard assay-cost cap. Randomized residual corrections support sequential testing despite inaccurate predictions. The analysis distinguishes identification, correction range, estimator variance, and decision efficiency, and permits a predeclared mixture of correction-model evidence on one assay stream. Controlled molecular simulations separate these mechanisms. On identical 320-assay observations under localized reversal, a betting test yields 97/100 positive decisions versus 22/100 for a classical empirical Bernstein bound at the same guaranteed one-sided error level. On disjoint learning histories under joint signal and calibration shift, fitted, centered, and mixed evidence yield 18%, 78%, and 76% positive decisions, respectively. The contribution is an evaluation protocol with matched diagnostic controls, not a new agent architecture or general inference method. Experiments use adaptive ridge campaigns and simulated outcomes, not an end-to-end agent or laboratory benchmark.