rodriguescarson/eligible-scroll-spiral-fits

Spiral fits for PHerc0191 and PHerc0826, published as unvalidated geometry

Four spiral-fitting runs on two prize-eligible scrolls, released as tifxyz meshes so that anyone can render them, run a detector on them, or measure them against their own fit. PHerc0191 has no other published mesh set.

These are surfaces produced, not coverage and not usable area. The winding density is unresolved and the audit below says so on the author's own numbers. Read that section before using the area figures for anything.

What is here

Per run: meshes/fitted/wNNN/{x,y,z}.tif plus meta.json per winding, the flattened samples, the render archives, satisfaction_metrics_fitted.json, and the fit checkpoint.

The three PHerc0826 windows overlap in z, so their figures do not add.

Provenance

• —PHerc0191 volume: 20250821151635-9.362um-1.2m-113keV-masked.zarr, shape 18977 x 8387 x 8387, voxel 9.362 um.
• —Umbilicus for PHerc0191: Drobkov's 64-point manual curve, commit a34224b76865, file sha256 b67a3aae154b9d27. PHerc0826 uses the official umbilicus.
• —Fitter: villa pinned at 757f70c0. Each run took 1.5 to 1.9 hours on one A40.
• —Prior art: gmDevi fitted the same PHerc0191 volume on Kaggle and published no meshes. This is not the first fit of that scan; it is the first published mesh set for it.

Winding density is unresolved, and the answer depends on the threshold

12 rays from the umbilicus at z = 12000, counting bright peaks in the raw scan between the innermost and outermost winding of the fit, at three peak-prominence thresholds.

Read both directions, because they disagree:

• —Against the strict peak count (k3) the fit has more windings than there are peaks on all 12 rays, 64-74 against 22-42. On that reading the fit is roughly two to three times over-packed.
• —Against the permissive count (k1) the fit is at or below the peak count on 4 of 12 rays (105, 225, 255, 285).
• —The autocorrelation period gives an independent sheet count of 28.6 to 76.9, which brackets our 64-74 from both sides depending on the ray.

So the honest statement is not "the fit is 2-3x over-packed". It is that sheet counting in this scan is threshold-dependent over a factor of three, and nobody has pinned the threshold. Our median winding pitch is 10.5 voxels against a scan peak pitch of 14-19. Treat every mesh here as unvalidated geometry until that is settled.

Two further negatives, published as they came out

Winding numbering does not transfer between overlapping windows. Across 400 shared slices, no winding of the second PHerc0826 fit matches a single winding of the first with 0.8 or more of its points; median share 0.28, offsets -6 to +18, while the surfaces themselves largely coincide at median 3.42 voxels. The test is confounded: consecutive windings sit a median 7.3 voxels apart with 31 of 89 pairs under 5 voxels.

A pre-registered A/B on winding count failed three of its four predictions. Arm B at 45 windings produced the same surfaces as arm A at 90: 45 of 45 matching at offset 0, share 1.0, median distance 0.00 voxels. shell_outer_winding_idx sets how far outward the fit exports, not winding density, so area follows extent (99.79 cm2 against 415.63). The fourth prediction, about render quality, is moot because the surfaces are identical.

Against pscamillo's published PHerc0826 meshes, measured one way only: 99.8, 99.7 and 90.9 percent of his mesh points lie within 20 voxels of one of our surfaces (median 3.7, 4.4, 4.5), though only 71, 58 and 23 percent lie within 5 voxels of the winding they mostly land on. The reverse distance was never measured.

Licence and credit

Meshes derive from Vesuvius Challenge open CT data and carry CC BY-NC 4.0. Scripts are MIT. The PHerc0191 umbilicus is Drobkov's work, used with the commit and hash recorded above. The fitter is villa, by the Scroll Prize team and contributors.

Analysis code and the JSON behind every number: https://github.com/rodriguescarson/eligible-scroll-atlas