One unit, and what it is not
A chain through the space, not a motif — but it moves with the experiment.
The difference between the two contexts is -0.811% [-1.520%, -0.243%], paired within animal and day. The interval excludes zero and a within-pair sign flip agrees. This is the only unit in the programme validated against the experimental design rather than against how it looks.
It is 361 segments across 46 of the 89 report animals, the largest single animal supplying 15.5% of them — so it is not one mouse's habit wearing a crowd's clothes.
What it is not
It is not one behaviour. It is a single-linkage chain: a typical pair of its members sits 0.398 apart against a linking distance of 0.190 — further apart than the threshold that built it. Single linkage joins A to C through B however far apart A and C are, so its members are not copies of one another and it must not be named as one behaviour.
What they share is a state, not a shape: they run at 0.272 of the speed of the same animals' other segments. A broad basin of near-immobility is what a chain looks like when the thing under it is genuinely wide — and it is why the token search had nothing to merge.
Freezing is a hypothesis, not a name. This is a fear-conditioning corpus, the unit is slow, and it is held more in the shocked context. Nothing here scores it against shock timing, and no rater has identified it. The clips below are the evidence a reader can check without taking any of that on trust.
Watch it, against a control that is not it
Each clip runs for its own segment's extent plus half a second either side — never a fixed window, because a unit shown in a longer clip is mostly not that unit. The control is drawn from the 98.1% of segments in no group at all, matched one-for-one on duration, so length cannot separate the two panels.
The skeleton is drawn on every clip and flagged frames would carry a red border. None do, in either panel: a segment is only selected when its abstain fraction is zero, and the abstain mask contains the bone-violation mask. The empty overlay is a property of the selection, not a finding about the island.
A third context exists and is not compared here. Context C occurs only on day 2 — both are exactly the same 479,962 frames — so an A-against-C contrast would measure the day rather than the arena. It was excluded by the registration, before any occupancy was computed.