Results explorer
13 runs · 6 held-out 1-hour streams · 10 Hz
y
level
event recall: fraction of ground-truth events with a boundary within ±3 frames
tokens per minute while the robot is idle vs active (log)
same 2-stage model at matched ~16×, with vs without the action input
All 13 runs · table (click a header to sort)
top = the level that feeds the main network · recall within ±3 frames · memory: cue→query gap < 2.5 min, n = 147, chance 25%