The label the method never saw

Fit non-negative CP at rank 3 to the 43 × 200 × 88 box of firing rates, then average each trial factor over the trials for each target.

• One component has almost the same trial weight whatever the target, and its time profile peaks early: something every trial does — the start of the movement, not its direction.
• The other two are strong for some targets and close to zero for others, disagree about which, and peak later. Between them they carry the direction.

Nothing in the fit used angle. The trial factor sorts the trials by a variable measured in the monkey’s behaviour because that variable is part of what generated the numbers. Be precise about what this is not: not a claim that a component “is” the 0° direction, but that a rank-3 summary of this tensor has to spend terms on something that varies with direction.