man5-surfaces

man5-surfaces documents traced surface geometry on PHerc. MAN5, a scroll with no published segments at all, for the Vesuvius Challenge. It exists because on a scroll with nothing published there is no reference surface to validate against, so the only checks available are internal consistency and an independent physical channel, and this repository runs both. Two headline numbers answer two different questions: 82.62 cm² of deduplicated surface passed the geometric quality gate, and 78.54 cm² of that is supported by the CT.

What it does not do

• It is not a ground truth measurement. On MAN5 there is no published segment to check against, so what the two channels give is agreement between checks that share no code, no thresholds and no inputs, which is weaker than a reference.
• Its gate has been tested on one scroll only.
• It does not claim an ink result. The pre-registered ink pass was withdrawn after two faults were found (the wrong preparation level, and no positive control), and the corrected re-run is not conclusive either.
• Its blind self-check is not a third-party check: the reader was the author, blind to the key but not independent of the experiment, and that reader also saw “lines” in the off-sheet controls, which the repository states makes an unqualified claim of lines here pareidolia.
• One of its own numbers, the 0.882 deduplication factor behind the 82.62 cm² headline, could not be recomputed from the archive: the pairwise overlaps it came from were not kept, and one contributing surface (B-100) is missing its tifxyz file entirely.

Reproduce

rederive.py needs neither the raw archive nor any dependency to check that the published figures are consistent with each other, and, given the archive, recomputes them from the audit records. public_areas.sh reads published segment areas anonymously over HTTP for the comparison scrolls used in the write-up.

How it works

A geometric gate accepted 86 surfaces totalling 93.6451 cm², counted by a sheetness threshold of 0.45 applied without exception: B-150, at 36.90 cm² and the largest single candidate, was rejected at sheetness 0.4251 with 21 of 49 tiles anomalous, against 4 of 49 for the largest surface that did pass. After deduplication the accepted set is 82.62 cm², built from A-150 (36.10 cm², sheetness 0.59), B-100 (14.90 cm², sheetness 0.62) and 84 smaller production patches.

A second, radiometric channel checks each surface against the CT independently: what fraction of its support falls inside the sheet mask within ±29 micrometres. Across the deduplicated set of 86 surfaces the area-weighted support is 95.07%, giving the 78.54 cm² figure. A-150 is 98.3% supported; B-100 is 78.53% supported, with roughly 3.2 of its 14.94 cm² unsupported at its east margin; B-150, rejected by the geometric gate, is only 2% supported by the CT. Across all 107 candidate surfaces, void fraction against the gate’s sheetness correlates at Spearman rho = -0.463 (p = 6.7 x 10^-7), with accepted surfaces averaging 1.8% void against 14.1% for rejected ones, the two measures derived independently of each other.

Re-voxelising the raw surfaces gives a deduplication factor of 0.9363 over 85 of the 86 surfaces, not the archived 0.882; the missing surface, B-100, is not stored in the archive, and the gross-area arithmetic (93.6451 minus 78.7021 equals 14.9430) matches B-100’s own area exactly, turning the gap into a testable prediction rather than an unresolved doubt: it implies about 40.2% of B-100 would need to overlap surfaces already counted, plausible since B-100 and A-150 come from the same growth series, but not itself measured. If 0.9363 held across all 86 surfaces the headline would read 87.68 cm², about 6.1% higher than published, so whatever this factor turns out to be, it does not inflate the total.

Investigation, implementation, measurement and documentation used Claude Code and OpenAI Codex agentically; PROVENANCE.md states, for every figure above, whether it was recomputed or taken from a work log.

man5-surfaces on GitHub, MIT licensed.