A single-author paper from Kim Phuc Tran, submitted to JMLR, challenges the common assumption that lowering predictive entropy is a meaningful safety mechanism. Using a noise-calibrated log-determinant functional, the author derives an exact spectral force formula and splits the learning dynamics into observable and prediction-null components. The central result is that entropy reduction may leave predictions completely unchanged or even increase risk for a protected group, depending on how constraints are specified.

The paper treats task, risk, and intervention constraints separately. For scalar feedback, it characterizes the full interval of coefficients that both preserve a required task-descent fraction and meet a requested entropy-rate target. Notably, that interval can become empty if interventions are limited to observable directions. For protected risks, the author provides a minimax feasibility certificate within a given subspace and norm budget, a robust extension for uncertain constraint rows, and a cone-projection formula for the entropy reduction remaining after protection.

The author is explicit about limits: the mathematics does not identify compression with alignment nor certify human well-being. Human authority enters only through specification and release policy. A memorization counterexample separates fresh-input information from sample-level information, and a noisy sample-code assumption yields a conditional generalization bound. Reproducible checks and a small real-data experiment show both feasible interventions and clear failures, reinforcing the paper's conclusion that entropy regulation should be characterized by its admissible regimes and limitations, not by universal claims of accuracy or safety.