PHerc0139: 0.22% voxel-size discrepancy between the two published scans; validated cross-scan transform available

[7jycwjmbfn-eng]

The two published scans of PHerc0139 (20250728140407, 9.362 um full scroll, and 20260413113053, 1.129 um mosaic ROI) have no published transform between them, and their nominal voxel sizes disagree by about 0.22 percent.

I registered the pair and validated the mapping on held-out data:

• Exhaustive global search (full scroll length x full 360 deg in-plane rotation), unique lock at z-score 27.2 over 49,920 candidate poses: rotation 183 deg, z offset 33.6 mm (gantry metadata independently predicts 35.5 mm).
• 3D block matching, 10,008 blocks, median NCC 0.906; robust rigid+scale fit.
• Split-half cross-validation: median held-out error 4.1 um, p95 12.4 um.

Two findings that may matter beyond this pair:

1. Voxel-size calibration: the fitted isotropic scale between the volumes is 1.00217 against nominal. Over the 40 mm ROI that is a ~130 um systematic error if the nominal sizes are used, i.e. more than three sheet thicknesses. Possibly an xray_magnification bookkeeping difference between the 2025-07 and 2025-12 reconstructions.
2. Residual deformation: after the rigid+scale fit there is a smooth residual field, 5.3 um median at 2 mm separation rising to 11.0 um at 16 mm, concentrated at the mosaic tile seams of the 1.129 um scan.

The transform in the project's transform.json schema (with 24 block-match landmark pairs embedded, so it can be re-fit from the file alone) is here, along with the code and the validation numbers:
https://github.com/7jycwjmbfn-eng/pherc0139-physical-audit

The same registration is used there to audit the published surface prediction volumes for this scroll against the 1.129 um scan as model-free ground truth (recall, side-of-sheet placement, miss bias), with the truth labels packaged as a dataset. Happy to adapt the transform format or landmarks if a different convention is preferred.

[7jycwjmbfn-eng]

Reopening with an update: the same registration pipeline has now been run on PHerc1203 (Grand Prize list; 20250820131727 at 9.362 um vs 20260319130212 at 2.403 um). That pair locks at z-score 44 with held-out error 2.4 um, and its fitted scale correction is only -0.025 percent. So the 0.22 percent discrepancy reported above for the PHerc0139 pair looks scan-campaign-specific rather than systematic.

Both transforms (the 0139 one in the transform.json schema with embedded landmark pairs for independent re-fitting), the audit built on them, and physical-truth label volumes for both scrolls are here: https://github.com/7jycwjmbfn-eng/pherc0139-physical-audit