Decoder stability
Decoder stability has two distinct meanings that are easy to conflate. The first is control-loop stability: with the user in the loop, an over-gained, laggy, or poorly tuned decoder can produce growing oscillations or runaway cursor motion, exactly as in any feedback system with insufficient margin. The second is longitudinal (day-to-day) stability: keeping a decoder's input–output mapping consistent across sessions so it works on day thirty as it did on day one, without a fresh calibration each time.
The two are linked by the user. Longitudinal stability matters not only for convenience but because a consistent mapping is what lets the user learn and consolidate a durable skill; a decoder that is recalibrated or that drifts each day forces the user to re-learn and never reaches expert control. A body of work therefore aims to stabilize the latent neural representation itself — aligning each day's population activity to a reference so that a fixed decoder remains valid despite turnover in which specific units are recorded — rather than repeatedly refitting the readout. Typical figures of merit are the number of days a decoder holds performance without recalibration and the consistency of the mapping over time.
Users consistently report preferring a slightly less accurate but stable decoder over a marginally better one that changes daily — predictability is itself a usability property, because it is what a human controller can build an internal model against.