We present a controller for supervising neural network training during architecture search that combines a trajectory-aware termination rule with a normalized thermodynamic stability metric. The stability metric is Phi = I × rho - alpha × Sₙorm, where I is identity normalized against the random baseline, rho is temporal coherence of the accuracy history, Sₙorm is confusion matrix entropy normalized for class count, and alpha = 0. 1. The trajectory-aware rule terminates training only when the best Phi achieved from the start of training remains below a viability threshold and improvement has stalled, distinguishing slow-starting viable architectures from genuinely non-viable ones. In a direct comparison on 30 architectures, the controller produced 0 false kills while patience-based early stopping produced 20 false kills (16. 7% kill precision) on the same architecture set. Across 660 total architectures spanning multilayer perceptrons and convolutional networks on 2-class, 10-class, and 100-class problems, the controller achieved 2 false kills (99. 7% kill precision). A threshold of 0. 25 applies to 2 through 10 class problems and 100-class multilayer perceptrons. A separately calibrated threshold of 0. 084 applies to 100-class convolutional networks. The controller is designed as a conservative kill filter: it accepts false passes to eliminate false kills, and its primary validated property is kill precision rather than balanced binary classification accuracy. The methods described in this paper are the subject of U. S. Provisional Patent Application No. 63/938, 279 (filed December 11, 2025). No license to implement or commercialize the described methods is granted by this publication. All rights reserved.
Shawn Barnicle (Fri,) studied this question.