Everywhere else on this site is one fit, examined after the fact. That's normal for a method that takes seconds to minutes per fit — you run the experiment, then you analyse it. GRIN takes about a millisecond. That's not just "the same workflow, faster" — it opens up things you can't do at all when fitting is slow: watching an estimate firm up trial by trial, catching drift while it's still happening, stopping a session the moment you're confident enough instead of running a fixed number of trials regardless.
Everything below is simulated, GRIN's stub (the trained network isn't wired in here yet — see the note in each section), and honestly labelled where the mechanism has a real limit.
Each layer is one checkpoint's fit; older ones fade as newer ones draw over them — the whole sequence is one picture of how the estimate moved, not just where it ended up.
A fixed-N protocol runs every participant for the same number of trials regardless of how the data look. Refitting every few trials means you can stop the moment you're confident enough — and skip the rest.