One pair of published ink predictions behaves unlike every other pair in the bucket: normal agreement at a loose budget, none at all at a tight one.
Segment: PHerc0500P2/segments/20250919184428-0500P2-wrap01_0919, predictions 20260417190342-new_canon_autoresearch_recipe and 20260709123958-mrg20736-1um-s1z2, both on volume 20250526151718.
Overlap of the top-q% of each map, as a multiple of the chance level q/(2−q) so the budgets are comparable. The baseline is every other checkpoint pair in the bucket, 55 of them:
|
q = 1% |
q = 5% |
q = 20% |
| other 55 pairs, median |
62.9x |
14.7x |
3.6x |
| other 55 pairs, minimum |
34.7x |
8.4x |
2.5x |
| this pair |
0.0x |
0.57x |
2.9x |
At 20% it is unremarkable. At 5% it is fifteen times weaker than the weakest other pair. At 1% the two top-k sets share not one pixel, where every other pair sits at 35x chance or better.
Not a rigid offset: an exhaustive search over every translation within ±60 px peaks at (−38, −34) with IoU@5% 0.0428, still an order of magnitude below the others.
What I cannot rule out from outside. This is a small map — 790,200 valid pixels, files of 148 KB and 240 KB, roughly six tile positions across at tile 256 / stride 128. #1374 and #1375 concern an overlap-add that imprints a ripple at the tile-grid frequency, which at this scale could plausibly decide a tight top-k. One map is also far more saturated than the other (5% threshold 226 against 196).
Reproduce with inkfloor, MIT, no GPU or credentials:
inkfloor corpus --kind model --samples PHerc0500P2
The full 56-pair run and its JSON are under results/ there.
Agent-assisted, as AGENTS.md anticipates.
One pair of published ink predictions behaves unlike every other pair in the bucket: normal agreement at a loose budget, none at all at a tight one.
Segment:
PHerc0500P2/segments/20250919184428-0500P2-wrap01_0919, predictions20260417190342-new_canon_autoresearch_recipeand20260709123958-mrg20736-1um-s1z2, both on volume20250526151718.Overlap of the top-q% of each map, as a multiple of the chance level
q/(2−q)so the budgets are comparable. The baseline is every other checkpoint pair in the bucket, 55 of them:At 20% it is unremarkable. At 5% it is fifteen times weaker than the weakest other pair. At 1% the two top-k sets share not one pixel, where every other pair sits at 35x chance or better.
Not a rigid offset: an exhaustive search over every translation within ±60 px peaks at (−38, −34) with IoU@5% 0.0428, still an order of magnitude below the others.
What I cannot rule out from outside. This is a small map — 790,200 valid pixels, files of 148 KB and 240 KB, roughly six tile positions across at tile 256 / stride 128. #1374 and #1375 concern an overlap-add that imprints a ripple at the tile-grid frequency, which at this scale could plausibly decide a tight top-k. One map is also far more saturated than the other (5% threshold 226 against 196).
Reproduce with inkfloor, MIT, no GPU or credentials:
The full 56-pair run and its JSON are under
results/there.Agent-assisted, as
AGENTS.mdanticipates.