Round-2 report — stroke & schedule ruled out

2026-07-11 · AI3D-338 · linebitmapsegmentation · program plan · round 1

Summary

Round 2 closed both open model-side levers, and both are nulls. (1) Per-class label stroke at physical paint width ({solid: 5, dashed: 3}): on both loss families, solid metrics reproduced round 1 exactly and dashed centerlines stayed found (clDice ~0.96–0.97) — the models simply painted the thinner label they were taught; only the measuring ruler moved. (2) 90-epoch schedule: early-stopped at 53 with test F1 0.8703 ≈ round-1's 0.8678 — the 60-cap was never binding. Together with round-1's data-volume null, the program has now confirmed from three independent directions that the remaining gap is label-geometry/data. One cheap implementation lever remains before declaring the exit: foreground-only class averaging in the dice/Tversky losses (background dilution), running as round 3.

Results (held-out 97-tile test)

Stroke runs — two rulers (native = s5d3 labels; yard4 = round-1 scalar-4 ruler)

run	ruler	F1 fg	F1 S/D	clDice	cov S/D	area S/D
`ft_r34_260624_s5d3` (best ep 24/35)	native	0.8390	0.858/0.820	0.961	0.920/0.913	1.14/1.10
`ft_r34_260624_s5d3` (best ep 24/35)	yard4	0.8291	0.858/0.801	0.969	0.920/0.694	1.14/0.73
`dice_focal_260624_s5d3` (best ep 35/46)	native	0.8529	0.865/0.841	0.962	0.883/0.858	1.04/1.04
`dice_focal_260624_s5d3` (best ep 35/46)	yard4	0.8042	0.865/0.743	0.967	0.883/0.603	1.04/0.62
Round-1 references (yard4 ruler): ft_r34 0.8678 · cov 0.916/0.938 — dice_focal 0.8750 · cov 0.875/0.889

The bold dashed numbers are geometry, not regression: a model taught 3 px dashes covers 3/5 of a 5 px ruler (0.73 and 0.62 area ≈ 3/5 with fit error). Solid columns are byte-identical to round 1 on both families — a clean negative control, exactly as predicted once the odd-width quantization discovery showed solids were already painted at physical width.

Schedule control

run	F1 fg	clDice	cov S/D	best ep / trained
`ft_r34_260624` (r1, cap 60)	0.8678	0.975	0.916/0.938	54 / 60 (no early stop)
`ft_r34_260624_e90`	0.8703	0.975	0.911/0.924	47 / 53 (early stop)

Verdicts

null Label-stroke realism is not the F1 ceiling lever. 2/2 loss families: width-invariant metrics flat, models faithfully track whatever stroke they're taught, thin dashes cost pixel-F1 headroom (a 3 px stroke tolerates ~1 px registration error far worse) and stop training earlier. Decision: keep scalar-4 (painted 5 px) training labels.
null Schedule is not a constraint. ΔF1 +0.0025 with 50% more epoch budget; round-1's "hit cap" was fluctuation.
win Small dashed-recall improvement on faint paint: both s5d3 runs catch the faint adjacent-carriageway dashes that round-1 dice_focal dropped — thin labels concentrate dashed gradient on the paint core. Not enough to offset the ruler mechanics, but informative for future label work.
status The program's converged operating point on this data: F1 ~0.87 · coverage ~0.92/0.94 · clDice ~0.975 (ft_r34 line, recall-first pick).

Round 3 — last implementation lever, then exit

Per Miro's priority (recall > precision), one genuine wiring flaw found during the dice-loss review is still untested: smp's multiclass Dice/Tversky average over all classes including background by default. Background dice saturates ~0.99 within epochs, so a third of the loss is a flat constant diluting the marking-class gradient — a plausible recall drag. smp exposes classes=[1, 2]; our registry doesn't pass it.

run	change	question
`ft_r34_260624_fg`	Tversky with `classes=[1,2]`	Does removing bg dilution buy real coverage on the recall anchor?
`dice_focal_260624_fg`	dice term with `classes=[1,2]`, focal unchanged	Does the calibration anchor recover FT-level coverage?

Registry change (~3 lines, expose classes) delegated to Grok 4.5 in Cursor. Scalar-4 labels, 90-epoch cap, seed 1337 — directly comparable to rounds 1–2. Exit rule: if coverage gains stay inside the seed band, the program declares data/label-geometry the bottleneck (three nulls + round-1 volume null) and closes with a final program report naming the ship checkpoint and the concrete data work (gore/wide-paint representation, faint-tile confirmation pass, asphalt-edge rule).