the voice control setup routes what i say based on where i'm looking, and for a while it was wrong in the dumbest ways. the first classifier only knew left and right — a single axis split — so when i looked down and to the right at the input box, my head pitch dragged the iris offset across the split and it called it the other window. pure horizontal logic can't see down. the fix is a 2d classifier: calibration records both x and y offsets for each anchor, and the current look goes to whichever anchor it's nearest.
then the gate was dropping everything unless i was looking at the assistant window, which is the design, but the paste function kept grabbing focus and pulling the assistant forward while i was looking elsewhere. the fix: paste only fires when the user is already looking at the target window, otherwise it buffers silently and shows the state in a draft panel.
the real progress came when i stopped asking for fixes and demanded debugging tools first. a replay tool that traces the gaze log with timing and confidence, a movable bench window with a dark background for testing, and visible state everywhere. running a real trace showed 58 percent of the time classified as one window, 41 as the other, flips about once a minute, and the longest continuous look running about a minute of real time.
calibration was the sneaky bug at the end. the anchor for the assistant window got captured too extreme because the prompt text drew at a pill in the corner while my eyes stayed on the ring, and the polluted split made a normal look land in the dead zone, sticky. verification now re-samples each anchor with the ring visible and rejects the capture if the new split misclassifies its own anchor, keeping the old good calibration. dead zone narrowed, tilt tolerance relaxed, and the suite went 51 for 51. garbage in, verified out.