Voynich | edkSG Research Volumes · EDKSG-VOY-V018 · Edition 1.0.0 · 5 September 2026, Singapore
Research status: executed local source-lineage audit of the five event relations that supply the f88r okol/qokol conditional comparison. Four relations retain the same analytical feature tuple across the named historical rows examined; one retains the same current target but splits on the predecessor’s final character. No image-level adjudication, model refitting, full-page consensus or decipherment is claimed.
The family is not uniformly unstable.
Across the five relations that create the scored okol/qokol comparison, four preserve the same combination of current-form outcome and predecessor-y status in the First Study Group, Don Latham and Takeshi Takahashi rows inspected here, and in the current ZL3b-derived analytical source.
The fifth relation also preserves its current target: okol. What changes is the immediately preceding form. The First Study Group and current ZL3b-derived source carry qockhol; Latham and Takahashi carry qockhoy. That one change flips the predictor used by the frozen model.
The correction is therefore concentrated: four comparator relations survive this local lineage check, while one decisive predecessor classification remains unresolved.
This does not make the remaining uncertainty small in consequence. Because the disputed relation is also the sole unprefixed current outcome in the five-case family, changing its predictor class moves the common frozen-model score from approximately +0.378 to −0.906.
1. Stability of the comparison, not a vote on the glyph
Volumes 15 through 17 established three separate facts. A documented l/y alternative exists. It reverses the local conditional score. The correction does not reverse every compatible result elsewhere.
A remaining ambiguity was whether the five-case family was itself a fragile construction assembled from many transcription disagreements. If several current forms, predecessor endings or word boundaries changed across the named rows, then focusing on one terminal stroke could understate the source problem.
This audit therefore freezes the statistical question and inspects only the five relations that supply it. For each relation we retain three analytical features: the current form’s okol/qokol family, whether the current form carries the initial q, and whether the immediately preceding represented span ends in y.
We do not ask which transcription is “best”. We do not count rows as independent votes. We ask which features required by this exact calculation agree and which do not.
2. The compared source records
The current analytical representation is the ZL3b-derived f88r diplomatic object used in the recent numbered volumes. Its relevant running-text loci are 19, 27 and 29.
The historical comparison comes from the Landini–Stolfi interlinear archive as exposed by the Voynich Information Browser. The relevant aligned rows are f88r.P2.7, f88r.P3.12 and f88r.P3.14 for the First Study Group, Don Latham and Takeshi Takahashi.
These records are not statistically independent observers. The interlinear archive converted and aligned earlier transliterations into a common representation, and the broader transliteration history contains inheritance and editorial relationships. Source plurality is useful for locating disagreement; it is not a licence to turn four files into four votes.
3. The five relations
| Relation | Source address | Current form | Named-row result |
|---|---|---|---|
| 1 | f88r.19 / P2.7 | okol | Split: qockhol → okol in FSG and current ZL; qockhoy → okol in Latham and Takahashi |
| 2 | f88r.27 / P3.12 | qokol | Stable: chol → qokol; predecessor non-y |
| 3 | f88r.27 / P3.12 | qokol | Stable: qokol → qokol; predecessor non-y |
| 4 | f88r.29 / P3.14 | qokol | Stable: chey → qokol; predecessor y |
| 5 | f88r.29 / P3.14 | qokol | Stable: chody → qokol; predecessor y |
All five current target forms required by this comparison are preserved across the rows examined. The four prefixed comparator relations also preserve the predecessor-y classification.
The disputed relation is different in exactly one way that matters to the model: the predecessor changes from a non-y ending to a y ending. Its current okol remains present.
4. Local feature stability is not textual consensus
The surrounding source lines are not identical. On P3.12, for example, the First Study Group begins the line with a form rendered differently from the Latham and Takahashi rows. On P3.14, the initial form also differs among rows. Other characters and segmentation details vary elsewhere on f88r.
Those differences do not enter the five relations listed above in a way that changes their current outcome or predecessor-y class. That is why this audit uses the phrase feature-tuple stability, not transcription agreement.
A later experiment asking a different question could be sensitive to those other differences. Stability is always claim-scoped.
5. One unstable relation can still dominate a tiny family
Four stable comparator relations out of five does not imply that 80% of the statistical conclusion is secure.
The disputed relation is structurally special. It supplies the only unprefixed current outcome in the family. The other four current outcomes are q-prefixed. Without that unprefixed occurrence, the family has no current-outcome variation and cannot provide the conditional comparison at all.
The uncertain predecessor ending is therefore attached to the observation with the greatest leverage over the local score. The source disagreement is concentrated, but so is the information.
This is why counting stable rows or stable relations cannot substitute for examining which role each observation plays in the calculation.
6. The frozen score remains conditional on that one classification
| Representation of relation 1 | Common-model gain vs zero | Hand-1 model gain vs zero |
|---|---|---|
| qockhol → okol | +0.378 | +0.272 |
| qockhoy → okol | −0.906 | −0.464 |
| Remove the disputed relation | Not scored | Not scored |
No coefficient is changed in this table. The four comparator relations are unchanged. Only the predecessor-y class attached to the sole unprefixed outcome changes between the first two rows.
The arithmetic therefore does not show broad instability of the whole family. It shows high leverage of one unresolved source classification inside a small matched set.
7. Why agreement is not counted as independent confirmation
The First Study Group, Latham, Takahashi, the interlinear archive and the current ZL effort belong to a documented history of transliteration, conversion, alignment and revision. Some source objects can preserve independent readings; others can inherit or transform earlier work.
For that reason, “three rows agree” is not automatically three independent observations of a glyph. The right record is a source graph: who read what, from which reproduction, under which alphabet, and how later files were produced.
The present result uses agreement only for a narrower purpose: to identify which analytical features are not presently known to vary across the named rows inspected. It does not convert that agreement into a probability that the underlying manuscript reading is correct.
8. The image verdict remains separate
Nothing in this lineage comparison resolves the terminal stroke on the physical manuscript. The high-resolution image task carried by Volumes 16 and 17 remains open.
If the relevant stroke cannot be distinguished at an available reproduction quality, the correct result is not observable at that resolution, not a forced l or y. A route failure to an image service is likewise an access state, not an observed mismatch.
A future image review should retain the entire local word, baseline, neighbouring strokes and preselected comparison examples. Cropping only the terminal can remove precisely the context needed to decide whether a stroke belongs to the candidate form.
9. What this volume changes
The local correction can now be described more precisely.
It is not presently supported to say that the whole five-case family dissolves across historical transcriptions. Four of the five analytical relations survive the named-row comparison in the features relevant to this model.
It is equally unsupported to say that the family is secure because most relations are stable. The one unresolved relation carries the sole unprefixed outcome and reverses the score when its predecessor classification changes.
The appropriate status is therefore: comparison structure largely retained; decisive predictor classification unresolved.
10. The next evidence job
The highest-value evidence remains the disputed physical reading or a documented editorial record explaining why a maintained transliteration chose one ending over the other.
A second bounded route is now also clear: extend this same claim-scoped lineage audit to the influential same-form families in the other fixed-model evaluation pages. That should be done before calling the larger positive total robust to transliteration uncertainty.
Such an audit should begin with the exact event relations contributing the most conditional score and ask whether their model-relevant features agree across source lineages. It should not perform a page-wide character-difference count and assume every difference matters equally.
Volume 18 shows that the local uncertainty is narrower than a wholesale transcription collapse and more consequential than a simple one-character footnote. One unresolved predictor sits at the hinge of the comparison.
Sources and reproducibility record
[1] Immediate predecessor. Vol. 17 — The Scope-of-Correction Check, EDKSG-VOY-V017. Its local l/y scenarios and correction-scope distinction are retained.
[2] Current analytical source. ZL3b-derived f88r diplomatic object, specifically running-text loci 19, 27 and 29.
[3] Named historical rows. Voynich Information Browser — f88r, exposing the First Study Group, Don Latham and Takeshi Takahashi rows from the Landini–Stolfi interlinear archive.
[4] Transliteration history. Voynich MS — Transliteration of the Text, used to preserve distinctions among historic transliteration efforts and later aligned/maintained files.
[5] Executed local record. EDKSG-VOY-V018-COMPARATOR-STABILITY-01. The package contains the five-relation lineage ledger and frozen-score checks. It does not contain a high-resolution manuscript image or claim an independent glyph verdict.
Collection: Voynich Research Library · Editorial framework: Wintour House.
EDKSG-VOY-V018 · Edition 1.0.0 · 5 September 2026 · Baseline EDKSG-VOY-B000. Claim-scoped source-lineage audit. No majority-vote adjudication, model refitting, independent palaeographic review or semantic decipherment is claimed.