Auditing Backward Transfer Against Learning Strength: A Reanalysis of Biologically Plausible Continual Learning

Change metrics can be misleading when compared methods differ in the reference point from which change is measured. We test this problem in continual learning, where a method can appear to forget less because it learned less before forgetting was measured. We examine ten papers on biologically plausible learning rules. Only two report how much the method and its baseline learned before forgetting was measured, and four provide no way to recover that quantity. We reanalyze five published comparisons after calibrating the relation between learning accuracy and backward transfer. Two retention advantages survive, one shrinks by about a quarter, one reflects re-learning rather than retention, and one reverses. The calibrated residual detects experience replay while not crediting two mechanisms expected to perform poorly in this setting. From this audit, we derive a claim--condition--control disclosure that records the conditioning variable, null relation, controls, operating range, and decision rule. Our convenience sample demonstrates an auditable failure mode and a practical correction but does not estimate its prevalence across fields.