# Pre-registration: the Voynich manuscript against medieval scribes

Written 2026-09-26 at about 15:45 UTC, before the full comparison set had been measured.

## What had already been seen

- Every Voynich number: the transliteration is ZL 3b (voynich.nu, sha256 bf5b6d4a…086096f), paragraph
  text only. Character entropy h2 (median over 20,000-symbol chunks) 2.03 bits with EVA letters as
  written, 2.19 with the benched glyphs as single symbols, 2.36 with the i-runs also single.
  Line-final information 0.089, line-initial 0.149 (EVA). Identical adjacent words 1.46 times the
  shuffled rate; one-edit neighbours 1.47 times.
- Three printed texts (Project Gutenberg 33849, 2000, 782).
- CATMuS Medieval, partially: while the text columns were still downloading, the measures were
  run on the 21 manuscripts with at least 5,000 words in the first ~23,000 rows [correction below: 74,774 rows]. All 21 were
  Castilian or French. Result: h2 3.06 to 3.27; line-final information 0.011 to 0.229, with the
  two highest (0.229, 0.219) French verse, where rhyme puts the same endings at the line's end;
  line-initial 0.006 to 0.101; identical-neighbour ratio 0.010 to 0.232; one-edit ratio 0.36 to 0.94.

## Predictions for every other CATMuS manuscript (at least 5,000 words, main text lines)

Same code (`engine.mjs`, `research/voynich-measured/run-manuscripts.mjs`), same thresholds:

1. None has h2 below 2.60 bits.
2. No **prose** manuscript has line-final information at or above 0.089. (Verse may: that is
   already known from the first 21.)
3. No manuscript, prose or verse, has line-initial information at or above 0.149.
4. No manuscript has identical adjacent words at or above the shuffled rate (ratio 1.0).
5. No manuscript has one-edit neighbours at or above the shuffled rate (ratio 1.0).

Each prediction is scored on the manuscripts not among the first 21, and reported whether it
holds or fails.

## Correction, same session, before the full data was measured

The partial run read **74,774** rows, not "~23,000" as written above: the file had grown further
than the progress log I glanced at. The saved record is `research/voynich-measured/results/seen-before-preregistration.json`
(row count and the 21 shelfmarks). Nothing else changes: those 21 are still the only manuscripts
whose numbers were seen, and a manuscript that had fewer than 5,000 words in those rows was not
measured and is scored with the rest.