The frontier in one sentence
You have finished Volume II. You can build a Kalman-filter or RNN decoder that hits state-of-the-art accuracy on a fresh calibration block. This volume opens with an uncomfortable fact: that same decoder may be measurably worse tomorrow, and much worse next month — not because you built it wrong, but because the brain and the electrodes beneath it will not hold still. Keeping a decoder accurate for years, without a daily calibration ritual, is the defining engineering problem of clinical BCI.
A decoder is a fixed function with parameters phi, fit on day-0 data. If the neural feature distribution p_t moves away from p_0, yesterday's phi is quietly solving the wrong problem today.
This is not a corner case. BrainGate and every other chronic intracortical program has had to confront it; research and commercial systems alike still lean on periodic recalibration. The question of this track is how far we can push toward a decoder that never asks the user to stop and recalibrate — and, just as important, where that push honestly runs out.
Three things that move under the decoder
It helps to separate the culprits, because they demand different fixes. First, the recording moves: micromotion, gliosis, and signal degradation mean units appear, fade, and change shape day to day. Second, the representation moves: even with a perfectly stable electrode, cortical tuning can slowly reorganise — representational drift. Third, the user moves: attention, arousal, strategy, and learning all shift the input.
On the day-to-week timescale of an intracortical array, much of what looks like 'drift' is actually recording instability — units the spike sorter gained or lost — rather than genuine cortical reorganisation. Getting the diagnosis right matters: you cannot fix an electrode problem with a fancier learning rule, and you cannot fix physiological drift by re-thresholding.
Why 'just recalibrate' is not the answer
The obvious fix is to recalibrate often. It works, and for years it was the only option. But daily supervised calibration is expensive in exactly the currency clinical users cannot spare: time, effort, and dignity. A person using a speech or cursor BCI does not want to spend the first quarter-hour of every day producing labelled training data. And supervised calibration presupposes the user can still produce clean, labelled attempts — not always true after progressive disease.
So the frontier reframes the goal: not 'recalibrate faster' but 'recalibrate without stopping, without labels, and eventually without the user noticing at all.' That reframing splits into self-recalibrating decoders and latent-space stabilisation (Guide 4), and continual learning that never forgets (Guide 5).
The shape of the rest of this track
Guide 2 anatomises drift and the one hopeful invariant beneath it — the stable manifold. Guide 3 covers the supervised and closed-loop workhorses (ReFIT, CLDA) and their ceiling. Guide 4 is the heart of the track: unsupervised, self-recalibrating, and latent-stabilised decoders. Guide 5 is the honest ledger — the stability–plasticity dilemma, the co-adaptation ceiling, benchmarks, and where the next decade points.