winding-ruler

Measurement tools and results for winding evidence in the Vesuvius Challenge spiral-fitting pipeline: how much human winding annotations move the fit, what the published lasagna signal can and cannot do, and the first collection-wide measurement of Herculaneum scroll winding geometry.

Full write-up: docs/SUBMISSION_winding_evidence.md

Headline results

• Human winding annotations are statistically redundant where verified patches are dense and worth +3–5 pp where patches thin out (PHerc Paris 4, two windows × two seeds, train-free evaluation).

• A calibrated estimator recovers human relative-winding labels at 93% for adjacent pairs — yet three progressively better constraint generators all degrade the fit. Root causes identified per generator.

  Revised 22 Jul 2026. The write-up attributed the residual to signal materialization resolution. That was pre-registered as testable and has since been falsified: regenerating the lasagna field at 2x finer scaledown does not improve winding accuracy (93.0% coarse vs 88.3% fine at dw=1 over 20 seeds; unit-step concordance 92.3% on both grids). The bottleneck is elsewhere.

• Winding pitch across the collection: a distribution, not a constant. Per-scroll medians center at 187.3 µm (IQR of medians 181.5–193.4; 35/36 scrolls between 160 and 210 µm), independent of scroll size and scan campaign — but pitch varies widely inside each scroll (Paris 4 alone spans 134–259 µm between quartiles), so no single number describes a scroll. The per-scroll λ table is the usable artefact; the scalar medians exist to cross-check methods against each other. Human-anchored count-based pitch ≈175–180 µm; the physical fundamental sits lower — ~145 µm on Paris 4 (see Resolved, 1 Aug below).

  Revised 22 Jul 2026. The original figures (207 µm, IQR 206–212, 34/35 within 190–242) were measured at pyramid level 2, which merges adjacent sheets: rerunning all 36 scrolls at level 1 lowers the pitch by 10.3% on average (36/36, sd 2.6) and drops the bias against the human anchor from +15.2% to +4.1%. On Paris 4 itself the level-1 estimate is 182.4 µm, +1.3% against the anchor. Current data is results/atlas_collection_v2.csv; atlas_collection.csv is kept as the v1 record. See atlas/build_atlas_v2.py for the derivation.

  Independent measurements, 30 Jul 2026. Aleksei Drobkov (AlexeyDrobkovStrikesBack on GitHub, alyalya1404 on Discord) first reported agreement using these atlas values as his anchor, caught the circularity himself, retracted it, and redid both scrolls independently: Paris 4 initially at ~180 µm/turn (Theil-Sen on the winding annotations; retracted by him on 31 Jul as unstable across voxel choices - his harmonic-safe remeasurement reads ~145 µm, under investigation here, see below) and PHerc1203 at 170-190 µm from radial layer counts on the BM18 volume, cross-checked by unrolled-length plausibility. The level-1 entries here read 182.4 and 192.3, consistent with both. He also supplied the structural note this debate was missing: one wrap is one papyrus sheet (~150-200 µm); the finer 10-40 µm structure is intra-sheet, so naive per-layer counts over-split delaminated sheets. Raw measurements to be linked here when he sends them.

  Resolved, 1 Aug: the discriminating test (mode vs median of adjacent-sheet gaps on this project's own GT) confirms his mechanism: mode 130-140 µm with secondary humps at the 2x/3x delamination harmonics. The values in this table are mixture medians of that distribution - reproducible periodicity in the surface predictions, internally consistent, but NOT physical fundamentals. His harmonic-safe methods place the Paris 4 fundamental near ~145 µm (CT autocorrelation 150, mesh mode 140, hand 143), and his five-scroll table (~145-230) confirms pitch genuinely varies, so the per-scroll table remains the right artefact. Every entry here needs the mode-vs-median treatment before being read as a fundamental; until then, read these values as what they are. Full record: constraint-gauge GATE0, addendum A26; his raw numbers and overlays archived there.

Values in the figure below are mixture medians of the sheet-gap distribution — reproducible periodicity in the surface predictions, not physical fundamentals (see Resolved, 1 Aug above).

Gate overlays — arms (a) vs (c) on the same W2 slice, near-identical to the eye; the measured +3–5 pp lives in sub-visual local deviations (pixel divergence 0.6–0.7%/slice, concentrated in the patch-sparse folds):

Layout

concordance/   pairwise Δw estimators v1–v1.5 + distance-confound test
generators/    constraint generators v1, v2, v2.1, v3 (all evaluated, all post-mortemed)
evaluation/    ensemble E1, difficulty D1, train-free evaluator recipe, run tags & seeds
atlas/         winding_atlas v1.2 — streaming collection-wide pitch measurement
results/       atlas_collection_v2.csv (current), atlas_collection.csv (v1),
               figures, per-run satisfaction logs
docs/          submission write-up + method notes

Reproducing

Everything runs against public data: the spiral-input PHercParis4 dataset (HF snapshot 2026-07-15, villa @ 37c37de) and the vesuvius-challenge open-data S3 bucket (streamed; no bulk download needed for the atlas). Single consumer GPU (RTX 5070, 12 GB) for the fit experiments; the atlas is CPU + network only. Fixed seeds throughout; pre-registered criteria noted inline in each script header.

Coordinate note: the dataset README states (z,y,x); point-collection JSONs are (x,y,z) in practice; track dbm arrays are (z,y,x).

Credit

Ray/pitch formulation follows Diego-dcv's technical note §2; per-scroll profiles complement iyando's stitched-instance profiles (PHerc 1218/0332); lasagna pipeline provenance confirmed by waldkauz; evaluator-shape guidance from sean (bruniss).

Tooling note

I work in Portuguese; English drafting and pair-programming were LLM-assisted, as in my previous projects. Measurements, experimental decisions, code, and all reported numbers are my own and reproducible from this repo.

License

MIT — see LICENSE.