Control mapping as a design variable: pricing the posterior-to-command step in motor-imagery BCI

A motor-imagery BCI converts a posterior over imagined movements into a command for a device, and that step is an argmax most reports leave unstated. We treat it as the independent variable. We decode posteriors from BCI Competition IV-2a once, cache them, and replay them through a simulated planar robot under four control mappings: argmax, posterior weighting, evidence accumulation, and shared control with an autonomous potential field. Nine experiments and 10,260 episodes cover nine subjects, three decoders, a controlled posterior-accuracy sweep, and four ablations. Switching from argmax to shared control buys as much closed-loop success as raising posterior accuracy by 4.5 points (95% CI 1.6 to 7.1), while the three decoders we test differ by 5.4 points of offline kappa in total. Shared control also flattens the dependence of success on offline kappa, from a slope of 6.04 to 4.14 (interaction -1.90, CI -2.84 to -0.97). Evidence accumulation moves both quantities the other way and falls to 0.10 success under a decoder whose confidence is understated, because it stops committing. A closed-loop result reported without its control mapping leaves an effect of this size unstated. Code and configurations: https://github.com/luminolous/mi-control-mapping (tag v1.0.0)