Part 2 showed r_norm stabilizes SBP features across days. Part 3 asks the next question — does a single nonlinear decoder trained on all 30 sessions, using both SBP r_norm and TCR z_norm, outperform the per-day Ridge baseline across every transfer horizon?
Part 1 showed that a static decoder trained on Day 0 raw SBP actively misleads by year three. Here's the preprocessing step that fixes it — no cross-day alignment, no labels, no retraining.
I loaded 30 sessions of chronic Utah-array data spanning three years and watched a trained decoder go from useful to actively harmful. Here's what the data showed me — and why I think it's a problem worth solving.
