When the Metric Cannot See the Model: Structural Decoupling in Simulation-Based Evaluation of Blockchain–AI Supply Chain Systems

Integrated blockchain and machine-learning systems for pharmaceutical supply chains are usually evaluated in simulation, and the headline result is typically a system-level detection rate reported beside the accuracies of the underlying models. This article examines whether such a metric actually responds to those models, using a twelve-node Hyperledger Fabric system with three analytic agents as a controlled test case. A one-factor-at-a-time ablation over three model improvements moved the reported counterfeit detection rate by exactly zero percentage points, because the metric was computed from severity-linked proxy signals rather than from model inference. We state a condition that detects this structural decoupling without running an experiment, show that recalibrating the composite score cannot repair it, and report a pre-registered redesign that replaces the proxies with live inference and with provenance derived from realized custody chains. The redesign restores coupling: degrading the classifier lowers system-level detection from 92.3% to 87.5%, with the same sign on all ten seeds. It also shows that blockchain provenance helps only when counterfeiting is concentrated at identifiable sources, that trust weights calibrated on proxies do not transfer, and that the corrected metric has its own blind spot: a 12.3% false-positive rate that five pre-registered repairs, each confirmed on held-out seeds, failed to reduce at the operating point. The protocol, code, per-seed records, and withdrawn conclusions are released, and a four-question checklist is derived for evaluators of such systems. Preprint, not peer reviewed. This work has been submitted to the IEEE for possible publication (IEEE Access, manuscript Access-2026-50542). Copyright may be transferred without notice, after which this version may no longer be accessible. Code, configuration, per-seed results and pre-registration documents: https://doi.org/10.5281/zenodo.21987357