In this work the authors audit two synthetic voltage‑clamp tasks to test whether a learned scientific controller can outperform a misspecified planner without using feedback. The experiments employ a training‑selected fixed protocol, current masking, a Bayesian reference model and an excluded target. Five recurrent policies repeat the same fixed protocol across 1,200 off‑grid rollouts; masking the current does not alter any decision. In a crossed four‑hypothesis task a fixed diagnostic pulse is followed by a single learned choice. The policies then rely only on the first measured current, yet their performance ties that of a correctly specified Bayesian controller; a fitted threshold model reproduces the learned actions without being that controller. The excluded‑target prediction gain is $4.6991\times10^{-6}$ nats per sample. The authors emphasize that feedback usage and protocol quality are separate tests, and these cases do not establish biological discovery or general reinforcement‑learning superiority.
Review