Cortical Decoders Trained on Other Individuals Age Slower Than Your Own

With adapting brain–computer interface technology, neural decoding algorithms are essential to understanding people's brains, and decoders must keep working for months — but recalibrating them per subject carries a high cost. A decoder fit to one animal's cortex drifts, losing accuracy over days, and the standard remedy is recalibration on fresh data from the same subject. We report the opposite regularity: decoders built from other individuals age more slowly than the subject's own. On the open BraiDyn-BC cohort (25 mice, 44 Allen-atlas cortical parcels, ~15-day operant protocol), a subject's own early-week decoder loses more AUC over two weeks than one pooled over the other 24 mice (+0.017 AUC, p = 5e-5, 73 of 100 mouse-event pairs). This is not a data-volume effect: at matched training-event counts, a decoder from a single other mouse still ages less (self minus one-other-mouse = +0.010 AUC). The effect survives nonlinear decoders (1-D CNN, GRU) and replicates on an independent two-photon dataset from a different laboratory (Allen Visual Behavior, 23 mice, VISp; +0.023 AUC, p = 0.003). We interpret this as personal decoders overfitting individual, drift-prone idiosyncrasies: the early-week cost of cross-individual decoding is repaid as late-week stability, suggesting population priors as an alternative to per-subject recalibration. A second contribution is the session-held-out protocol under which all of this is measured: only the within-subject arm can share a recording session with its test data, so a trial-level split inflates that arm alone.