Voynich | edkSG Research Volumes · EDKSG-VOY-V023 · Edition 1.0.0 · Prepared 6 September 2026, Singapore
Research status: completed targeted comparison of fifteen previously selected f105r relations against a displayed historical transcription attributed to Takeshi Takahashi. Two model-relevant features differ. This is a local source-reading audit, not an image verdict, complete alternate-transcription analysis, independent replication or decipherment. Publication is pending.
The next check had to include the evidence against our model, not just the evidence supporting it.
Volumes 20–22 identified four high-leverage families on f105r: the two strongest positive contributors and the two strongest negative contributors. We have now traced all fifteen of their admitted relations back to the retained ZL3b-derived source and compared the corresponding passages with a publicly displayed, Takahashi-attributed transcription. The target list was inherited, not selected to make this new comparison look favourable. [1–4]
Thirteen of the fifteen relations retain the same current-form family, prefix outcome and predecessor-y classification in the compared display. Two do not—and both changes affect the supportive families. One moves a current form from okeedy to oteedy. The other changes a predecessor from sheokedy to sheoked, removing its final y. [3–5]
That distinction is more informative than an overall percentage of matching characters. A small number of disagreements can affect precisely the observations carrying a model’s positive margin. Meanwhile, the two adverse families retain their relevant binary arrangements in this comparison.
The bounded result is a source qualification, not a corrected reading: the selected positive support is sensitive to attested alternatives, while the selected counterevidence cannot simply be dismissed as absent from the compared historical display.
1. From influence ranking to source inspection
Volume 22 established that the retained f105r page score stays positive after removing any single family, but has little margin after the two strongest positive families are removed together. That was an influence calculation on one representation. It did not establish that those families were more accurately transcribed than the others. [1]
Today’s question is different: when the actual source passages are compared, do the features needed by those influential families remain the same?
We kept the balanced target list from Volume 20: okeedy/qokeedy, okedy/qokedy, okaiin/qokaiin and okal/qokal. In the retained model, the first two contribute positively and the last two negatively. Together they contain fifteen admitted relations. [2]
Keeping both directions in the audit matters. Finding a doubtful supportive reading does not justify ignoring an equally doubtful adverse one. Conversely, a negative model contribution is not a reason to classify a source observation as an error.
This investigation is retrospective. We knew the influence ranking and earlier results before inspecting these passages. No historical blindness, external preregistration or independent reviewer is claimed.
2. What was compared
The present baseline is the repository’s ZL3b-derived f105r diplomatic object. The connector returned its text and object identifier. Eleven selected raw-field lines were transferred locally, and the fifteen inherited event offsets were checked against those lines. The complete repository file was not reconstructed and hashed locally in this run; a connector identifier is not represented as a newly completed whole-file verification. [3, 6]
The comparator is a University of Adelaide research wiki page explicitly identifying its text as a transcription by Takeshi Takahashi. The page has a fixed revision reference, 17231, and presents the text as space-separated words. The corresponding passages are also visible in the historical interlinear display linked from the Voynich folio catalogue. [4, 5]
A displayed transcription is not the same object as a complete raw historical transcription file. The wiki display does not preserve every confidence marker, alignment symbol or editorial annotation. It can establish that a displayed sequence differs; it cannot automatically tell us whether that sequence would meet every rule of our modern raw-input parser.
The comparison is therefore between source representations at identified passages. It is not a certified run of the unchanged parser on a complete independent transcription. We also do not assume that a mirrored historical text is the transcriber’s latest revision.
Direct downloads did not yield the complete origin corpus. The large interlinear repository file was also not acquired as a usable complete local source: one route returned empty content, and another failed UTF-8 decoding. These access outcomes are recorded, not converted into successful acquisition claims.
3. What counts as agreement here
For each selected occurrence, we compare three things: its current-form family after removing an optional leading q, whether the current form actually carries that q, and whether the immediately preceding represented form ends in y.
These are analytical features. They are not decoded roots, grammatical categories or statements about pronunciation. EVA’s role as a representation rather than a translation remains essential to the interpretation. [7]
Exact strings can differ without changing this feature description. At f105r.33, the retained split view supplies aiiin before qokaiin; the historical display supplies oraiiin. Both predecessors are non-y, and the current outcome is q-prefixed in the same family. That relation is feature-stable but not string-identical.
Other questions could be sensitive to the difference. A model using predecessor identity or actual word boundaries could not treat these two forms as interchangeable. Our agreement count is restricted to the features used by this one experiment.
The resulting inventory is twelve exact pair matches, one pair with changed spelling but unchanged model features, one changed predictor and one reassigned current-form family. Together these account for all fifteen selected occurrences.
4. A current form moves to another family
At modern source locus f105r.16, the retained raw field has oiir followed by okeedy. The compared historical display instead has oiir followed by oteedy. The corresponding historical interlinear passage also visibly contains these alternatives. [3–5]
The predecessor class remains non-y, and the current form remains unprefixed. But the exact current-form family changes. An occurrence contributing to okeedy/qokeedy in the baseline no longer belongs to that family in the alternative.
This matters because the model does not pool every word beginning o into one matched set. It conditions on exact current-form families. A change from k to t can therefore alter the relevant comparison even while leaving the binary prefix outcome unchanged.
We preserve the alternative oteedy event in a separate destination record. It is not deleted from the source and not silently reclassified as missing. Its receiving family would need to be rebuilt in a whole-page alternative analysis. That receiving-family calculation is outside today’s fixed four-family audit.
5. A predecessor loses its final y
At f105r.17, the retained representation has sheokedy immediately before qokedy. The Takahashi-attributed display has sheoked before the same current form. [3, 4]
This time the current-form family and its q-prefixed outcome stay fixed. The predictor changes: the first represented predecessor ends in y and the second does not.
The difference therefore has another route into the calculation. It changes how the observed current outcomes align with the predecessor feature, rather than moving the current event to another exact-form group.
Neither difference identifies the correct manuscript reading. A maintained modern source may incorporate a justified revision; a historical reading may preserve an ambiguity worth checking. We did not obtain an editorial rationale or inspect a sufficiently detailed manuscript image for either disputed form.
6. The selected adverse arrangements remain
The two negative families receive the same scrutiny. Their five selected relations preserve the family and binary-feature arrangements in the compared display, including the adverse pairing of a y-ending predecessor with an unprefixed outcome.
The three-case okaiin/qokaiin family still has its two q-prefixed outcomes after non-y predecessors and its unprefixed outcome after a y-ending predecessor. The exact predecessor at locus 33 changes as described above, but not its non-y classification.
The two-case okal/qokal family likewise retains its opposed arrangement. This source comparison provides no basis for discarding that counterevidence merely because a positive coefficient fits it poorly.
There is a qualification. The historical display’s spaces do not recover the raw uncertainty status of every boundary. This result is agreement of the displayed model-relevant arrangement, not proof that all five cases would enter a complete historical-source run under an identical confidence-marker policy.
7. A fixed-family sensitivity calculation
To quantify the identified differences, we retain the common coefficient at 1.5440411135795047 and evaluate the same four preselected family names. We change only the mapped readings, without fitting any new parameter. The current form that becomes oteedy is retained outside this four-family projection.
| Focal family | Retained cases | Cases remaining in that family | Retained contribution | Mapped-reading contribution |
|---|---|---|---|---|
| okeedy / qokeedy | 5 | 4 | +1.420 | +0.891 |
| okedy / qokedy | 5 | 5 | +0.992 | +0.638 |
| okaiin / qokaiin | 3 | 3 | −1.240 | −1.240 |
| okal / qokal | 2 | 2 | −1.044 | −1.044 |
| Focal subtotal | 15 | 14 | +0.127 | −0.755 |
The calculation distinguishes two effects. The okeedy family loses one non-y, unprefixed member and its positive contribution falls. The okedy family retains five members but changes one predictor classification, also reducing its contribution. The two adverse contributions stay unchanged under the inspected mapping.
The four-family subtotal consequently changes sign. That is a useful local sensitivity finding, but it must not be promoted into a reversal of f105r’s whole-page result.
8. Why this is not a corrected page score
A full alternative-source analysis would have to find every relevant occurrence under the other representation, apply an explicitly compatible segmentation and uncertainty policy, reconstruct all affected families and account for new as well as lost events.
Today’s cohort was chosen from the retained ZL-derived source. It contains the fifteen previously influential occurrences. It is not an exhaustive search for every occurrence that a historical reading might add to these families elsewhere on the page.
In addition, the reassigned oteedy occurrence can affect a receiving family outside the quartet. Simply subtracting its lost contribution from the original page score would ignore that movement. There may also be other differences in the historical page that this focal comparison has not assessed.
The local sign change therefore means that the selected balance is reading-sensitive. It does not mean that the complete page has been shown to reverse, or that the old transcription has defeated the new one.
This is why the artifact records both the out-of-family event and the limited scoring population. A changed denominator is part of the result, not a footnote to remove before presenting the total.
9. What was actually checked
The retained fifteen events are selected programmatically from the parent event index using the four family hashes. Each baseline predecessor and current string is recovered from the source offsets and checked against the manually transferred current raw-field line. Every comparison pair must occur exactly once in its corresponding historical display line.
The score implementation uses exact integer binomial coefficients inside a numerically stable normalization. A second check enumerates all fixed-outcome allocations for each of the eight small baseline and alternative family tables and agrees to the stated tolerance.
Twelve unit and regression tests passed. They cover the fifteen-case inventory, thirteen unchanged feature descriptions, three changed exact pairs, preservation of the reassigned event, unchanged adverse tables, the fixed coefficient, invalid inputs and the distinction between an unavailable score and zero. These are computational tests, not twelve manuscript experiments. [6]
The manually transferred source lines remain an input limitation. The program can detect a mismatch with the inherited offsets or missing comparison phrase, but it cannot independently certify that every character was copied from the display correctly. The research package records source URLs, revision information and the scope of the transfer.
No external reviewer, image adjudication, complete historical-parser run or independent transcription vote is claimed.
10. What the next stage should resolve
The high-leverage audit has now found two different source risks in the supportive evidence: family reassignment and predictor reclassification. It has also checked the adverse families rather than assuming they are less reliable because they oppose the model.
The next source work should seek the raw historical records and a documented reading rationale, then verify the disputed forms against a provenance-bound image. A whole-page rerun would need a carefully stated cross-source boundary policy, not a silent substitution of displayed spaces for raw confidence markers.
The archive’s earlier primary observations remain unchanged. The alternate readings are scenarios over known material, not new manuscript evidence to append to the historical total. The coefficients also remain frozen.
The model has not gained a meaning. The source question has gained two exact addresses and a clearer accounting of what can change. That is the advance recorded by this balanced reading check.
Sources and research record
[1] Volume 22 — The Support-Concentration Check, published 6 September 2026, post 154888. The prepared edition retains its 5 September edition date.
[2] Volume 20 — The Counterevidence-Retention Check and Volume 13 — The Frozen-Coefficient Check, defining the balanced focal families and preserved analytical inputs.
[3] ZL3b-derived f105r diplomatic object, retrieved through the connector on 6 September 2026. Returned Git blob identity: e98b4423762a3c3fdcb96426f5b210b56c9f6b02. Selected raw lines were transferred; the entire blob was not newly reconstructed locally.
[4] University of Adelaide research wiki: f105r, revision 17231. The page attributes its displayed transcription to Takeshi Takahashi. Consulted 6 September 2026; displayed word spacing is not treated as a complete raw annotation record.
[5] Historical f105r interlinear display, linked by the Quire 20 catalogue. Used to check that the differing passages appear in the historical representation record; browser parsing did not preserve every locator or row attribution, so no new named-row census is inferred from it.
[6] EDKSG-VOY-V023-BALANCED-READING-01. The accompanying package contains the parent-event binding, selected source-line transfers, complete fifteen-case mapping, out-of-family record, executable score calculation and twelve-test log. This is an executed local comparison with bounded source access, not a fully automated source ingestion.
[7] René Zandbergen, Transliteration of the Text, especially the distinction between representation and translation and the relationships among historical and maintained files. Consulted 6 September 2026.
Catalogue continuity: Volumes 3A and 3B retain their distinct published titles and original URLs. Volume 23 follows the actually published Volume 22; earlier volumes and the hub are unchanged.
Editorial framework: Wintour House. EDKSG-VOY-V023 · Edition 1.0.0 · Prepared 6 September 2026. No independent review, whole-page reversal, source correction or decipherment is claimed.