A Context-Conditional Audit of Trial-Pairing-Dependent Neural Gain in Motor-Cortex Decoding

ObjectiveNeural-only decoding performance does not identify how much prediction neural history adds beyond task structure or recent output. We tested how this increment changes when context and temporal information boundaries are made explicit. ApproachWe defined context-conditional neural gain as held-out error reduction from adding context-residualized neural history to a context model. Fully nested, whole-trial cross-fitting excluded predicted trials from nuisance preprocessing, fitting, and selection. Trial replacement tested reliance on correct neural-behavioral pairing. Matched-seed calibration and linear and nonlinear sensitivities accompanied two Neural Latents Benchmark datasets, nine paired LINK dates (18 development sessions), and 20 prespecified untouched LINK sessions from the same macaque. Main resultsIn LINK, fixed neural-only R2 near 0.31 coexisted with gains of 0.3012/0.3125 beyond phase, 0.0193/0.0209 beyond geometry by phase, and 0.0077/0.0098 beyond that context plus measured-output history for center-out/random-target tasks. Fully nested analysis retained positive means for two neural feature definitions; replacement made all four negative. Geometry-conditional spiking-band-power gain reduced context-model mean squared error by 4.92%. In the untouched sample, geometry-conditional gain was positive in 20/20 sessions (mean 0.02158), contracted to 0.00755 with measured-output history, and exceeded replacement in 20/20. MC_Maze showed gain 0.00051 beyond a strong template, whereas MC_RTT retained 0.06846 beyond lag-matched cursor history. Three prespecified nonlinear-neural random-feature maps preserved this contrast while changing magnitudes. SignificanceUnder the tested specifications, neural gain depended on context, information availability, and model family, and the residual correction depended on correct trial pairing. Neural-decoding reports should therefore state the outcome, context, temporal boundary, model class, grouped validation, and pairing control. The audit is predictive and model-relative, not causal.