Voynich | edkSG Research Volumes · EDKSG-VOY-V025 · Edition 1.0.0 · 7 September 2026, Singapore
Research status: executed full-page alternate-source reconstruction of f105r using the Takeshi Takahashi H rows preserved in the complete Landini–Stolfi IVTFF interlinear archive. The frozen coefficients and parser lineage are unchanged. This is an alternate transcription test, not a new manuscript observation, image adjudication, independent replication or decipherment.
We can now run the whole page.
The complete Landini–Stolfi interlinear archive explicitly identifies transcriber code H as Takeshi Takahashi, describes his contribution as a complete Voynich transcription, and preserves the f105r H rows with definite and dubious word breaks, alignment fillers, drawing interruptions and editor-defined paragraph breaks. [3]
After converting only the archive’s declared alignment fillers into absence—not manuscript characters—and reconstructing the editor-marked paragraph starts, the unchanged preservation and boundary pipeline accepts all 37 f105r loci and recovers 10 paragraph starts. Two isolated title loci remain outside the marked-paragraph experiment, as in the retained geometry model.
The Takahashi page yields 139 admitted oR/qoR transitions, 11 informative exact-form strata and 36 informative observations. With the common coefficient still frozen at the value learned before Volume 13’s target acquisition, its conditional gain is approximately +1.273.
The retained ZL3b-derived f105r representation scored approximately +2.649 under the corresponding primary view. The alternate-source whole-page result is therefore lower by approximately 1.376, retaining about 48.1% of that positive gain.
The structural clue survives a complete Takahashi-page reconstruction with its sign intact, but its strength is materially smaller. The result is neither a reversal nor a source-invariant confirmation.
1. The missing source object is no longer missing
Earlier volumes repeatedly encountered a source-access limitation. Individual diplomatic page files were available, and normalized historical displays were available, but a complete raw alternate transcription suitable for a page-level rerun had not been placed inside the same executable pipeline.
The public voynich-public repository exposes LSI_ivtff_0d.txt, a 1,754,295-byte interlinear archive. Its own header states that it was derived from the Landini interlinear file and later additions, and that Takeshi Takahashi’s complete transcription was incorporated after being fetched on 25 November 1998. The archive assigns code H to Takahashi. [3]
This distinction matters. We are no longer inferring Takahashi’s line from a space-normalized web page. We have a source object whose format documents which characters are transcription content, which are alignment fillers, and how line and paragraph breaks are represented.
2. What was converted—and what was not
The existing eduKate parser accepts a bounded merged-EVA profile rather than the full historical interlinear syntax. We therefore construct a deterministic H-only wrapper around the thirty-seven f105r H rows.
Two archive conventions require explicit handling.
First, alignment fillers. The archive documentation says ! and % are interlinear alignment fillers. They do not encode Voynich characters. The H-only wrapper removes those fillers while preserving the surrounding text, definite periods, dubious commas, drawings and inline comments.
Second, paragraphs. The archive says line-final = denotes a paragraph break and explains that the interlinear editors inserted these breaks by examining reproductions of the manuscript. In the connector-exposed representation these paragraph ends appear as the same end markers consumed by our parser. We restore a corresponding start marker at the first P0 locus after each paragraph end.
This produces exactly ten marked paragraph starts, matching the retained f105r analytical frame. We do not use the ZL word strings to choose those starts.
3. One metadata choice is held constant deliberately
The historical archive’s page header records Currier hand Y, while the retained analytical frame used by the frozen model assigns effective hand 3 to f105r.
For the common model this difference is irrelevant because the same coefficient applies to every hand. For the previously frozen partial-sharing model, however, hand determines which deviation is used.
We therefore retain hand 3 as the analytical metadata for the score comparison. This keeps the model comparison focused on the alternate transcription rather than silently changing the text and the model-routing label together.
That is an experimental control, not a claim that the historical archive’s hand label is wrong.
4. The alternate page produces more admitted transitions but slightly less matched information
| Representation | Admitted transitions | Informative strata | Informative observations | Frozen common gain |
|---|---|---|---|---|
| Retained ZL3b-derived f105r | 133 | 12 | 37 | +2.649 |
| Takahashi H reconstruction | 139 | 11 | 36 | +1.273 |
The alternate page has six more admitted transitions overall but one fewer informative stratum and one fewer informative observation. This is a useful reminder that corpus size and comparison opportunity are different quantities.
The model scores exact current-form families. A transcription can add represented transitions while simultaneously moving observations among families or removing the within-family variation needed for this conditional test.
5. The positive sign survives, while the magnitude contracts
The retained common-model score is approximately +2.648786. The full Takahashi reconstruction scores approximately +1.272781. The difference is −1.376006.
Expressed as a ratio of positive conditional gain, the Takahashi reconstruction retains about 48.1% of the retained-source result.
This is a materially weaker result. It is not close enough to justify saying that transcription choice has little effect.
It is also still positive. The two reading differences isolated in Volumes 23 and 24 were not sufficient to predict the full-page direction on their own, because other source differences elsewhere on the page also enter the exact-form inventory.
The complete-source rerun therefore resolves a question those local projections could not: under the H transcription, the page remains directionally compatible with the frozen common clue, but with substantially reduced support.
6. The hand-specific refinement does not become a clear winner
The frozen partial-sharing model scores approximately +1.351 on the H reconstruction, slightly above the common model’s +1.273.
That small difference does not overturn Volume 12’s broader finding that hand-specific refinements had not earned a robust preference. This is one already-exposed page, one alternate transcription and the same retained hand-routing label.
We record the ordering because suppressing it would be outcome-selective. We do not promote it into a new model-selection conclusion.
7. The three doubtful-gap policies collapse to one result here
For this H-only reconstruction, the split, join and exclude-touching procedures produce the same admitted event count and the same frozen scores.
This does not prove that word-boundary uncertainty has disappeared from the manuscript. It means the H rows used here do not express the relevant candidate boundaries through the same dubious-comma pattern that caused policy divergence in the retained ZL3b-derived page.
A source can therefore change not only the characters but also which sensitivity axis is available to test. That is another reason not to interpret three identical policy scores as three independent confirmations.
8. What happened to the Volume 24 bounded estimate?
Volume 24’s bounded propagation gave approximately +1.926 after carrying the two already-identified reading changes through their affected and receiving families while holding all other retained strata fixed.
The full-page H reconstruction gives approximately +1.273. The difference shows that those two high-leverage changes were not the whole transcription delta.
Other H-versus-ZL differences alter event eligibility, exact-form membership and predecessor classifications elsewhere on f105r. A local patch can be valuable for causal accounting, but it cannot substitute for a complete alternate-source rerun once the complete source becomes available.
The earlier bounded result remains valid for the question it asked. Volume 25 supersedes it only as an estimate of the full-page H representation.
9. This still is not independent manuscript replication
The Takahashi transcription and the retained ZL-derived representation both refer to the same physical page of the same manuscript. They are alternative representations of one historical object, not independent experimental samples.
The interlinear archive also has an editorial history. It aligns multiple transcriptions and documents editorial insertion of line and paragraph breaks. Its source header explicitly describes those transformations. [3]
The alternate-source result therefore measures sensitivity to a documented transcription lineage. It does not multiply the number of manuscripts, scribes or independently sampled historical observations.
10. Verification completed
The reconstruction contains all thirty-seven H rows for f105r. The wrapper produces thirty-seven preserved records, ten marked paragraph starts and two unmarked isolated title loci. Every preserved payload round-trips through the unchanged parser.
Eight regression checks pass, including the exact event count, informative denominators, frozen common score, frozen partial score and equality of the three downstream gap-policy results.
The accompanying executable script reruns the parser, extractor and both frozen models from the saved H-only wrapper.
The H-row subset was transferred from the connector-exposed complete archive and checked against the public f105r interlinear display. This is not claimed as a fresh image transcription or an independently keyed diplomatic edition.
11. What we can now say more confidently
The predecessor-y / q-prefix relationship on f105r is not solely an artefact of the retained ZL3b transcription. A full-page reconstruction using Takahashi’s H rows leaves the frozen common conditional score positive.
At the same time, the score is substantially smaller. The relationship is therefore sensitive to transcription lineage in magnitude and composition.
This is stronger evidence than a two-character sensitivity patch because the entire page is reconstructed. It is weaker than a prospective independent test because the page, question and model are all development-exposed and the two transcriptions describe the same manuscript object.
Nothing in the result assigns semantic meaning to q, y, okeedy, qokedy or any other represented form.
12. The next research threshold
With a complete H-page rerun now possible, the same procedure can be applied prospectively to other high-influence pages before inspecting their H-based score. That would be a cleaner test of transcription-lineage sensitivity than choosing pages after seeing where the current model performs best.
A second route remains image-level adjudication of high-leverage disagreements. That would supply a different kind of evidence: not another transcription file, but a direct assessment of whether the represented alternatives are supported by the manuscript image.
The most important design rule is now clear: source comparison must be whole-page when possible, preserve alternate observations rather than deleting them, and keep model coefficients frozen while the representation changes.
Volume 25 moves the pathway from local transcription sensitivity to complete alternate-source evaluation. The clue survives, but it becomes smaller—and therefore more precisely bounded.
Sources and reproducibility record
[1] Volume 24 — The Receiving-Family Check, EDKSG-VOY-V024.
[2] Volume 13 — The Frozen-Coefficient Check, supplying the frozen models and retained f105r baseline.
[3] Landini–Stolfi interlinear archive LSI_ivtff_0d.txt, Git blob f5e722e709874749b9b4f04b30406c7fb8e67fcc, 1,754,295 bytes, exposed by the public xenoglyph-ai/voynich-public repository. Its header identifies H as Takeshi Takahashi, documents filler semantics, and defines definite, dubious, line and paragraph breaks.
[4] Historical f105r interlinear display, used as a visible secondary check on the transferred f105r H passages.
[5] EDKSG-VOY-V025-FULL-PAGE-ALTERNATE-SOURCE-01. The package contains the H-row ledger, deterministic wrapper, inherited parser/model code, event and stratum outputs, eight regression checks and an executable rerun script.
Collection: Voynich Research Library · Editorial framework: Wintour House.
EDKSG-VOY-V025 · Edition 1.0.0 · 7 September 2026. Full-page alternate-source sensitivity audit. No independent manuscript replication, image adjudication, source correction or decipherment is claimed.