A prospective protocol for reliable closed-loop deployment of computational models

Computational models are increasingly embedded in closed-loopscientific workflows where predictions feed fixed downstream rules.Under fixed interfaces, average accuracy can poorly guide reliabilitybecause these rules selectively amplify decision-changing errors.We introduce Prediction-Target Validity (\textbf{PTV}), a zero-leakage protocolthat nominates the prediction target to preserve before decisivedeployment outcomes are observed. In a frozen molecular-screening replaywith a fixed top-$K$ stability acquisition rule (2,416 candidates,$K=200$, 15 rounds, 23 surrogates), the lowest validation-MAE surrogateis among the weakest deployment performers in the pool (42.1\%false-stable rate), whereas thePTV-nominated stability-threshold target approaches the within-libraryoracle (6.1\% versus 2.1\%).Across 60 held-out fixed-interface systems, PTV reaches 88.3\% top-1agreement with the within-library target-level oracle and 0.047 meannormalized regret. PTV provides anauditable validation protocol for computational-science workflows inwhich surrogate predictions feed fixed acquisition, screening, thresholdor control rules.