17 gate-passed questions · 12 cells · 8 occupied · go to this type's cells
Bins were cut on this corpus at 240×135@16 (CLEAR_PX 25, motion floor 19 px). Margin bins: a near-edge band [1, 1.25) above the gate boundary, then corpus quantiles; salience bins: a visibility floor, then quantiles.
landing gap (frames) / 3; 'same': one bin · measurable on 15 questions
| margin bin | numeric range |
|---|---|
| near-edge [1,1.25) | [−∞, 1.25) |
| bin0 | [1.25, ∞) |
| state:same | (state / single bin) |
Ask: is this the same quantity the gate certifies on for this type? Are the states (e.g. 'same', 'gone') the right one-bin exceptions? Are the edges where a change in difficulty actually happens?
min queried-object peak area (frac of frame) · floor 0.0007716
| salience bin | numeric range |
|---|---|
| <floor (0.0008) | [−∞, 0.0007716) |
| bin0 | [0.0007716, 0.008827) |
| bin1 | [0.008827, 0.01404) |
| bin2 | [0.01404, ∞) |
Ask: does this axis capture what makes a question of this type perceptually hard, independently of the answer? Is the floor the right anti-gaming boundary?
branch 2 spread default: n_airborne_ever (cv) · labels: bin2
Compatible margin bins per context label (cells outside these are definitionally impossible):
| answer | n |
|---|---|
| neither lands before the video ends | 3 |
| the green cube | 2 |
| the green cylinder | 2 |
| the orange ball | 2 |
| the yellow ball | 2 |
| they land at the same time | 2 |
| the red-and-green striped cube | 1 |
| the blue cone | 1 |
| the red-and-purple striped cube | 1 |
| the blue-and-yellow striped cylinder | 1 |