Measured Critical Periods: Freeze Timing Produces a Nonmonotone Accuracy Profile

Critical-period claims in artificial networks usually compare two conditions. We sweep freeze timing at every epoch and report the whole profile, including the interior optimum and its seed spread. Freeze epoch is swept exhaustively over the training run; the never-freeze and always-frozen endpoints are included as controls. The claim holds in part: 3 of the 4 predicates the experiment computed in advance hold, so we report the split rather than an overall verdict, and identify which parts of the claim the data establish. Principal quantities: profile range over seed sd = 39.044 seeds needed for interior optimum = 13; best freeze epoch = 7, with intervals where shown being 95% half-widths over the repetitions run.