Date: 2026-10-28
- Layout layer (N-001).
inductive/layout_layer.pyscreens minimal line-edge alternations on half E: one glyph written or dropped at the start of a line or paragraph, one glyph in place of an ending at the end of a line. Kept at three times the inside share, 15 stems, z ≥ 5: 46 rules. Paragraph start p-, t-, k-, f-; line start s-, y-, d-, t-, o-, p-; line end -m for -r, -l, -in, -iin, -n, -ir, -dy, -y, -g for -r, -l, -d for -r. Controls: plain Latin in the same lines gives no rules; Latin with three planted rules gives all three back, restores 75 % of planted words exactly and rewrites 0.01 % of the rest. - Registered before half C was read (
out/prereg_S007.yaml, SHA-256 d7bfd7d8…): H-G-LINEEDGE, H-G-LAYOUT-ROBUST, H-G-LAYOUT-MASK, and the crowding test of H-G-COPIED. - Results on half C (
t_layout_s007.py): 42 of 46 rules replicate; normalising 7 % of words removes 68 % of the line-position information (dropping the edge words removes more). Glyph entropy, word-order gain and the absence of repeated four-word phrases are unchanged under both treatments. Form-function coupling is already inside the language range on raw half C (0.086 against 0.095), so H-G-LAYOUT-ROBUST fails as registered; on the whole book it stays above every language (0.149 raw, 0.124 normalised, 0.113 interior). H-G-LAYOUT-MASK partly supported (coupling only). Crowding: -m 33 % in full lines against 28 % in short ones, z 2.6, below the bar: inconclusive. - Exploratory: on the normalised text the battery's topics follow Currier language and scribe less (NMI 0.53 to 0.43, 0.54 to 0.42 over five seeds) and section as much as before; every scribe uses the line-edge layer, B more than A. The homophonic solver on the normalised text: no key transfers (best gap 0.32 against 0.34 raw).
- Q-001 had only the forgery answer; added "a fifteenth-century original" (T-09, H-TH-ORIGINAL) and scored both on the same eight findings.
- Literature audit: two searches (history, object and pictures; text, methods and data) returned 266 items; after removing 52 already held, 214 were added to
sources/manifest.yaml(cited_in: lit-audit:S-007), each with the questions it bears on. Every DOI and arXiv id was checked against Crossref or arXiv from here; 8 remain citation-only (verified: false). Four bad citations corrected: "Hodgins 2016" (now Hodgins 2009 letter report, Hodgins 2012 talk, Zyats et al. 2016), "Stomeo & Golob 2011" (does not exist), "Clemens 2016 manuscript description" (not a chapter), "Davis 2017" (now Zhang 2017 and Davis 2019). Book chapters 1, 2 and 17 now cite the real sources. - Question audit: eleven questions added (Q-007, Q-008, Q-016, Q-017, Q-018, Q-024, Q-034 to Q-037, Q-043) and thirteen next steps (N-011 to N-023), each naming the method and the data it needs. Data the audit could not find in open digital form: Latin herbal and balneological texts (Circa instans, Tractatus de herbis, De balneis Puteolanis), Lingua Ignota and Enochian lexicons, a Voynich glyph segmentation with ground truth, a labelled dataset of medieval plant drawings.
- Site: theories that ACH ranks "strongly disfavoured" now show as disputed, not ruled out (a ranking from one rater is not a refutation).
- Open: N-011 Naibbe cipher; N-012 ACH sensitivity; N-014 why the line-edge layer (CATMuS, ReF); a second rater.