Each row is one clip. Left to right is where the detector scored it on a key or note it never practised on; the line in the middle is the cutoff. When the two colours sit on opposite sides of the line, the model's scratch paper carries the difference.
Audio Flamingo has the same shape: 28 brain layers of 3,584 numbers. No single number means anything on its own; meaning lives in patterns across many numbers at once, the way a chord is several notes together.
Two real clips from the roughness test: the same note (C), same average loudness, different flutter. Switch between them and watch what changes at each stage.
Each cell is one number, averaged over the clip. Warm = high, cool = low.
All 144 clips at the same layer, one row each; this clip is ringed. The detector squeezes its 4,096 numbers into one left-right score.
The per-slice readings were not saved, so stages 3 and 4 show each clip’s average over its roughly 75 slices. The two grids look similar to the eye; the difference is a faint pattern spread across thousands of cells, which is exactly what the detector is built to find.
The left-right position is the simplest possible detector: average one group, average the other, draw the line between them. Each clip is scored by a detector built without its key, so a split here is real hearing, not memory.
Up and down carries no measurement. Each row is simply one clip: the top group first, the other below, each sorted from low notes to high. The bar from the centre line shows how far, and which way, the detector scored that clip.
Scores are spread so the cutoff sits at the centre line. Major vs minor shows the wrong-way lean: many top-group bars point the wrong way.