Accuracy retention and width allocation at fixed parameter count: A reproducibility correction to The φ-Compression Law

The January 2026 φ-Compression Law preprint attributed an efficiency-retention product near one to golden-ratio layer widths. We show that its product is exactly normalized accuracy for every width schedule: GR = A/A0. Thus neither a sequence-specific conservation mechanism nor an upper efficiency bound e follows. An executable audit finds 162 configurations in the released code rather than the claimed 216, mismatched transformer parameter claims, and seeding after model construction. A prospectively frozen comparison trains nine SimpleCNN profiles with exactly 57,824 parameters and one larger reference, each for fifty epochs under three seeds. The validation-selected non-Lucas comparator is [43, 56, 52]. Mean test accuracy is 75.940% for Lucas and 76.417% for this comparator. The paired mean difference is -0.477 percentage points, with seed differences ranging from -5.340 to +5.340. The comparison measures finite performance under matched parameters and epochs, not matched compute or a number-theoretic mechanism. The historical conservation, e-optimality and transformer extrapolations are withdrawn. Source, protocols, model outputs and independent analysis make the narrower claim inspectable.

The reproducibility archive contains the relevant pinned and recovered source, frozen protocol and selection receipt, all 30 final model checkpoints and 1,500 eligible epoch records, split indices, canonical test IDs and labels, saved test predictions and logits, independent numerical checks, checkpoint replay outputs, tables, figures, and generation code. Raw CIFAR images and software environments are excluded. The versioned verification correction and interrupted attempt are preserved with their qualifications.

This is a corrective new version of the January 2026 preprint, DOI 10.5281/zenodo.18315993. It is a preprint and has not undergone external peer review.