Voynich | edkSG Research Volumes · EDKSG-VOY-V027 · Edition 1.0.0 · 7 September 2026, Singapore
Research status: executed the third predeclared whole-page alternate-source check from the Volume 13 evaluation set. f34r was selected in Volume 26 before its Takeshi Takahashi H score was inspected. The same frozen model and exact-form conditional rule are used. This is not independent manuscript replication, image adjudication, semantic interpretation or decipherment.
The third page remains positive.
Under the retained ZL-derived representation, f34r supplied only one informative exact-form family and two informative observations. Its frozen common-model gain was approximately +0.500.
The Takahashi H reconstruction supplies 31 admitted transitions, two informative exact-form strata and four informative observations. Its frozen common-model gain is approximately +0.999.
The increase is not a diffuse page-wide shift. The retained okedy/qokedy comparison survives. A second perfectly aligned family, oky/qoky, becomes informative in the H representation. Each contributes essentially the same small conditional gain.
The f34r result therefore remains directionally compatible with the predecessor-y/q-prefix clue and becomes stronger under the Takahashi representation, but the evidence is still only four informative observations across two tiny families.
1. Why this page matters despite its small score
f34r was the smallest contributor among the three Volume 13 target pages. That is precisely why it had to be included after f105r and f77r. An alternate-source programme that tests only the largest favourable pages would bias its own evidence base.
The page was fixed as the next target in Volume 26 before the H score was inspected. That is a cleaner decision boundary than choosing it because its alternate transcription happens to help or hurt the model.
The prospectivity remains bounded. The page itself, its retained score and its earlier source metadata were already known. The test is prospective only with respect to the uninspected H-based outcome.
2. The Takahashi text is structurally different from the retained page
The historical f34r display identifies fifteen Takeshi Takahashi paragraph-text rows. The page is herbal and contains repeated interruptions by the illustrated plant. Those interruptions matter because the analysis never joins text across a drawing.
For the executable reconstruction, plant interruptions are therefore represented as drawing barriers rather than deleted. This preserves the rule already used throughout the numbered pathway: an excluded or interrupted region must not cause its neighbours to become artificially adjacent.
The reconstruction retains two paragraph blocks. It does not import semantic information from the plant image, and the drawing itself contributes no text token to the statistical model.
3. More admitted events, but still very little direct information
| Representation | Admitted transitions | Informative strata | Informative observations | Frozen common gain |
|---|---|---|---|---|
| Retained ZL-derived f34r | 29 | 1 | 2 | +0.500 |
| Takahashi H reconstruction | 31 | 2 | 4 | +0.999 |
The page gains two admitted transitions overall, but the more important change is that one additional exact-form family now has the required outcome and predecessor variation.
Even so, four informative observations are still sparse. The page cannot be described as a broad replication of the model. It supplies two small same-form comparisons.
4. The retained okedy/qokedy family survives
The original informative family is okedy/qokedy. Under the Takahashi reconstruction it contains the same model-relevant arrangement:
- qotol → okedy: predecessor does not end in y; current form is unprefixed.
- cholchedy → qokedy: predecessor ends in y; current form is q-prefixed.
With one observation in each predictor and outcome class, the positive frozen coefficient gives the observed alignment more conditional probability than the zero reference.
This is the same structural direction as in the retained page. The alternate transcription does not erase the one comparison that originally made f34r scoreable.
5. A second family becomes informative
The Takahashi page also contains an oky/qoky family with the same two-case alignment:
- shotchy → qoky: predecessor ends in y; current form is q-prefixed.
- al → oky: predecessor does not end in y; current form is unprefixed.
This second family was not an informative matched stratum in the retained f34r evaluation index.
Its appearance explains the score increase. The page has not become twice as informative in a general sense; it has gained one additional two-case exact-form comparison whose arrangement matches the frozen positive coefficient.
6. The arithmetic is almost exactly additive
Each two-case perfectly aligned family contributes approximately +0.499625 at the frozen common coefficient. The Takahashi page contains two such informative families.
okedy / qokedy ≈ +0.499625
oky / qoky ≈ +0.499625
--------------------------------
f34r H total ≈ +0.999250That simple decomposition is valuable because it prevents the page total from sounding more complex than it is.
No optimization, new penalty or target refitting is required to produce the increase.
7. The hand-specific frozen model is also positive
With the same retained hand-2 analytical routing used for the earlier fixed models, the frozen partial-sharing procedure scores approximately +1.151 on the H reconstruction.
That is larger than the common score on this page. As in Volumes 25 and 26, the ordering is recorded rather than used to reopen model selection. These alternate-source pages are already development-exposed and are not a new model-selection dataset.
8. The three boundary policies again coincide
Split, join and exclude-touching treatments produce the same admitted-event count and the same common-model score on this H reconstruction.
This does not mean f34r has no uncertain boundaries. It means the particular H representation and reconstructed drawing barriers do not create the same doubtful-seam alternatives that caused policy divergence in the retained ZL-derived pages.
Identical sensitivity results remain related views of one source representation, not separate confirmations.
9. The three alternate-source page movements are now complete
The original Volume 13 target set has now been reconstructed page by page under the Takahashi H lineage:
- f105r: positive but weaker, from about +2.649 to +1.273.
- f77r: positive and stronger, from about +2.545 to +4.046.
- f34r: positive and stronger, from about +0.500 to +0.999.
The direction of the source-lineage change is therefore not uniform. One page weakens; two strengthen.
That is a more informative result than saying one transcription is generally “better” for the model. Source representation changes exact-form membership, comparison opportunity and predictor alignment in page-specific ways.
10. What the third result adds to the research pathway
Before these alternate-source volumes, a reasonable concern was that the fixed-model transfer result might depend heavily on one maintained transcription lineage.
We can now make a narrower and better-supported statement. All three original target pages retain a positive common-model score when reconstructed from Takahashi H text under the inherited scoring rule. However, the magnitude of that score is not source-invariant.
This is evidence for persistence of the structural direction across two transcription lineages on these pages. It is not independent manuscript replication, because both transcriptions describe the same physical folios.
11. The next aggregate must not erase the page movements
With all three H pages now available, the next natural calculation is the three-page alternate-source aggregate.
That aggregate should not be published as a single number without decomposition. Volumes 20 and 21 already demonstrated how stable totals can conceal large opposite movements. The same control applies here.
The aggregate must therefore report each page’s retained score, H score and change, the family-level drivers of the largest movements, and the total only after those components are visible.
12. Bounded conclusion
The prospectively selected f34r Takahashi reconstruction remains positive and roughly doubles its small frozen common score, because a second perfectly aligned two-case family becomes informative. Together with Volumes 25 and 26, all three original fixed-model target pages now remain positive under the alternate transcription lineage, while their magnitudes change substantially and in different directions.
No semantic value is assigned to q, y, okedy, oky or any other represented form.
Sources and reproducibility record
[1] Volume 26 — The Prospective Alternate-Page Check, which fixed f34r as the next page before H-score inspection.
[2] Volume 13 — The Frozen-Coefficient Check, supplying the frozen common model and retained f34r baseline.
[3] Voynich Information Browser — f34r, exposing the Takeshi Takahashi H paragraph-text rows and page metadata.
[4] Landini–Stolfi interlinear archive conventions used in Volumes 25–27. Plant interruptions are retained as drawing barriers in the executable wrapper rather than deleted.
[5] EDKSG-VOY-V027-THIRD-ALTERNATE-PAGE-01. The accompanying package contains the result ledger, exact page accounting and ten regression checks.
Collection: Voynich Research Library · Editorial framework: Wintour House.
EDKSG-VOY-V027 · Edition 1.0.0 · 7 September 2026. Third alternate-page sensitivity audit. No independent manuscript replication, image adjudication, source correction or decipherment is claimed.