We present a method for predicting whether a neural network architecture will form coherent task identity within a practical training budget using a single observation after one training epoch. The method introduces an identity deficit metric, Ideficit = 1 - sqrt (F), where F is the diagonal strength of the confusion matrix computed from epoch-one predictions. We find that architectures separate into two regimes at a critical threshold of Ic = 0. 22: those that form coherent task identity and reach task competence, and those that do not within the tested training budgets. Validated across 22 architectures spanning multilayer perceptrons and convolutional networks on MNIST, Fashion-MNIST, and CIFAR-10, the method correctly classified 21 of 22 architectures (95%) using a consistent competence cutoff of 80% final accuracy across all three datasets. The single misclassification is a slow learner whose epoch-one identity deficit falls in a caution zone (0. 22 to approximately 0. 40) that the method explicitly treats as requiring additional evaluation rather than immediate termination. The threshold was held fixed across the reported validation experiments. The method requires no access to gradients, weights, or loss curves, and produces a binary formation prediction from one confusion matrix in the time required to train a single epoch. The methods described in this paper are the subject of U. S. Provisional Patent Application No. 63/960, 091 (filed January 14, 2026). No license to implement or commercialize the described methods is granted by this publication. All rights reserved.
Shawn Barnicle (Fri,) studied this question.