Voynich | edkSG Vol. 19 — The Boundary-Lineage Check

Voynich | edkSG Research Volumes · EDKSG-VOY-V019 · Edition 1.0.0 · 5 September 2026, Singapore

Research status: executed claim-scoped source-lineage audit of the dominant ol/qol comparison on f77r. Seven of nine model-relevant relations remain directly comparable across the named transcription rows examined; two depend on word-break decisions that differ across those records. No model refitting, image verdict, full-page consensus or decipherment is claimed.

Most of the f77r score comes from one family.

Under the fixed common coefficient used in Volume 13, f77r contributes approximately +2.545 conditional log-likelihood units against zero. A single exact current-form family, ol/qol, contributes approximately +2.472 of that amount—about 97.1% of the page’s positive score.

That concentration makes the family the correct place to test source robustness.

The result is mixed in a useful way. Seven of the nine event relations retain the same model-relevant current-form family and predecessor-y class across the Currier, First Study Group, Takeshi Takahashi and John Grove rows examined here and the current ZL3b-derived analytical source. Two relations do not satisfy that cross-lineage condition because the source records disagree about a word break after qol.

Those two are not silently thrown away. Under a conservative sensitivity that keeps only the seven relations with shared segmentation and feature classification, the dominant-stratum gain falls from +2.472 to approximately +2.069. If every other f77r stratum is held at its previously calculated value, the page remains positive at approximately +2.141. [2–5]

The f77r result therefore depends on transcription segmentation for its exact magnitude, but its dominant positive comparison does not disappear when the two lineage-sensitive relations are removed.

1. Why audit the largest contribution first?

A source uncertainty should be propagated according to leverage, not according to how visually interesting it seems. Volume 18 did that locally for f88r: one disputed predecessor sat at the hinge of the only scoreable family. The same principle can be applied to a positive result.

Volume 13’s frozen common model scored f77r at +2.5448665230. The page contains four informative exact-form strata under the primary split policy. Their contributions are not equal. The ol/qol family contributes +2.4722023926 by itself. The remaining informative structure contributes only about +0.073 in net combination.

That makes the dominant family the first source-lineage target. If its event construction depended on a fragile or idiosyncratic word-break convention, the page headline could be substantially overstated. If the core relations survived across different transcription records, that would narrow the source concern without certifying the manuscript reading.

The audit therefore starts from the statistical dependency graph and works backward to the representation. It does not scan the page for arbitrary character disagreements and assume that all differences matter equally.

2. The original nine-relation family

The original stratum contains nine events. Six have a predecessor ending in y and a current q-prefixed form; three have a predecessor not ending in y and a current unprefixed form. In conditional-count notation this is:

n = 9
r = 6   # predecessor-y observations
m = 6   # q-prefixed current outcomes
k = 6   # overlap of predecessor-y with q-prefixed outcome

At the frozen common coefficient β ≈ 1.544041, this arrangement contributes +2.472202 against the zero-coefficient reference.

The nine source relations, expressed in the current ZL3b-derived representation, are:

Relation	Modern locus	Predecessor	Current form	Feature tuple
1	f77r.13	qokeedy	qol	y → qo
2	f77r.18	sheckhedy	qol	y → qo
3	f77r.32	cheol	ol	non-y → o
4	f77r.33	chedy	qol	y → qo
5	f77r.38	chcphey	qol	y → qo
6	f77r.38	shedy	qol	y → qo
7	f77r.40	daiiin	ol	non-y → o
8	f77r.44	qepchedy	qol	y → qo
9	f77r.45	ain	ol	non-y → o

3. The named historical comparison rows

The Voynich Information Browser exposes the f77r interlinear archive with separate rows for Currier, the First Study Group, Takeshi Takahashi and John Grove. The modern ZL3b-derived source is a later maintained representation and is compared separately.

The named rows are valuable because they preserve both agreement and disagreement. They must not be counted as independent votes: the interlinear archive aligned and converted earlier transcription work, and later transliterations belong to a documented history of source reuse and editorial revision.

Our question is more modest. For each of the nine event relations, does the named row preserve a separately represented ol/qol current form and the same predecessor-y class?

4. Seven relations retain the model-relevant structure

Seven relations meet that criterion across all of the named rows checked.

At P10, sheckhedy → qol is preserved. At P24, cheol → ol is preserved. At P30, both chcphey → qol and shedy → qol are preserved. At P32, daiiin → ol is preserved. At P37, ain → ol is preserved.

P36 contains an exact predecessor spelling difference: Currier has a form rendered as qopchedy where the other compared rows use qepchedy. That difference matters to a full transcription but not to this binary feature: every version ends in y, and every version retains the following qol. For this claim, the feature tuple is stable.

This is an important discipline. “Same model feature” is not the same as “same source text”. A source-lineage audit should report both.

5. Two relations depend on word-boundary decisions

The two exceptions occur at P5 and P25.

At P5, the current ZL3b-derived source records a doubtful separator after qol: effectively qokeedy.qol,sheedy. Under the primary split policy used by Volume 13, that doubtful seam creates a separate qol analytical span. John Grove’s row also separates qol and sheedy with a definite boundary.

But the Currier, First Study Group and Takahashi aligned rows do not present a separate word break there in the normalized display; the sequence appears as qolsheedy. The underlying interlinear uses alignment filler rather than a Voynich word-break character at that point.

At P25, the current ZL3b-derived source similarly uses a doubtful separator in chedy.qol,keedy. First Study Group, Takahashi and Grove give a separate qol. Currier instead presents qolkeedy without the same split.

These are not tiny typographic differences for this experiment. They determine whether qol exists as the same current-form family at all.

6. The segmentation policy cannot be separated from source lineage

Earlier volumes treated doubtful gaps as an explicit analytical choice: split them, join them, or exclude pairs touching them. That work held one source representation fixed.

The present audit adds another layer. Different transcription rows may disagree on whether there is a word break to classify in the first place. A comma in one source, a dot in another, and no break in a third are not three values of one downstream policy. They are different source representations entering the policy.

A robust pipeline must therefore record:

• which transcription row supplied the text;
• what boundary that row represents;
• which analytical boundary policy was then applied;
• which current-form family resulted;
• and whether a comparison survives across source-lineage alternatives.

Collapsing those layers would make a model appear insensitive to segmentation merely because the disagreement was removed before the sensitivity analysis began.

7. A conservative common-segmentation intersection

For a bounded sensitivity check, we retain only the seven relations whose model-relevant structure survives across Currier, First Study Group, Takahashi, Grove and the current ZL3b-derived representation.

This is not a claim that the excluded two are wrong. It asks a different question: what remains if we require the comparison to stand on relations that the checked lineages all permit us to represent in the same ol/qol family with the same predecessor-y class?

The retained counts become:

n = 7
r = 4
m = 4
k = 4

At the unchanged coefficient, the conditional gain is +2.068814.

That is lower than the original +2.472202, as it should be: two favourable y→qo relations have been removed. The positive comparison nevertheless remains substantial.

If the other f77r strata are left exactly as previously calculated and only this dominant stratum is replaced by the conservative intersection, the page-level gain is approximately +2.141478.

This adjusted page number is a local sensitivity accounting, not a full alternate-transcription rerun. The other f77r strata have not yet undergone the same lineage audit. We report the calculation precisely so that it cannot quietly become a stronger claim later.

8. What changes—and what does not

The exact magnitude changes materially. The dominant family loses about 0.403 log-likelihood units under the conservative intersection, roughly a sixteen-percent reduction in that stratum’s original gain.

The sign does not change. More importantly, the result does not depend exclusively on the two segmentation-sensitive events.

This contrasts with f88r, where the sole unprefixed outcome and its predecessor reading sat at the hinge of the only scoreable family. On f77r, the positive alignment is repeated across seven cross-lineage-comparable relations, including four y→qo and three non-y→o cases.

The comparison is still sparse and dependent. Several observations lie on the same page, some source rows share ancestry, and repeated forms do not become independent trials because they appear at different loci.

9. Why this is not a transcription-consensus result

The phrase “seven stable relations” needs a boundary.

We did not compare every character in every relevant line. We did not adjudicate every word break on f77r. We did not verify the interlinear rows against the manuscript image. We did not establish that Currier, FSG, Takahashi and Grove are independent observations.

We compared only the features necessary for one exact current-form stratum. Line-level differences outside those relations remain visible in the source records. A future model using another character, a longer context window or another segmentation rule could be sensitive to them.

This is claim-scoped reproducibility: enough source comparison to evaluate the lineage sensitivity of one statistical contribution, and no more authority claimed than that comparison supplies.

10. The source-risk profile of f77r is now clearer

Before this audit, the f77r result was positive under three gap policies within the current ZL3b-derived representation. We now know more about a different axis of uncertainty.

The dominant positive family is not wholly created by two modern doubtful seams. Seven of its nine relations retain the required model structure across the named historical rows examined. The two doubtful-seam relations do, however, increase the original magnitude and should not be treated as source-independent confirmations.

The strongest wording we can justify is therefore:

The dominant f77r ol/qol association remains positive under a conservative intersection of the model-relevant relations shared across the checked transcription lineages, while its exact magnitude is sensitive to two disputed word-boundary representations.

That is stronger than saying the result is “robust”, and weaker than saying the transcription uncertainty destroys it.

11. The next audit target

f105r is now the obvious next target. Its positive page score is similar in magnitude to f77r’s, but its contribution is distributed across more informative strata and includes important negative strata. That structure makes a source-lineage audit more complex and more informative.

The appropriate procedure is to rank the f105r stratum contributions by absolute influence, map their event relations back to the source, and audit the model-relevant features across available transcription rows. A result should preserve both positive and negative influential families.

A full parent-source acquisition would make this work more systematic. It remains unfinished. Piecemeal page files are useful for bounded audits, not a substitute for a verified complete corpus.

The source question has moved from “does another transcription differ?” to “which differences can actually change the claim we are making?” Volume 19 records one case where that distinction matters.

Sources and reproducibility record

[1] Immediate predecessor. Vol. 18 — The Comparator-Stability Check, EDKSG-VOY-V018.

[2] Fixed-model evaluation source. Vol. 13 — The Frozen-Coefficient Check, EDKSG-VOY-V013. Its frozen common coefficient, f77r event index and page score are retained without refitting.

[3] Current f77r analytical representation. ZL3b-derived f77r diplomatic file, including the relevant loci 13, 18, 32, 33, 38, 40, 44 and 45.

[4] Named historical rows. Voynich Information Browser — f77r, exposing Currier, First Study Group, Takeshi Takahashi and John Grove rows from the Landini–Stolfi interlinear archive.

[5] Interlinear source and filler semantics. text16e6.evt, J. Stolfi’s interlinear archive release, used to distinguish alignment filler from represented word-break characters.

[6] Transliteration history. Voynich MS — Transliteration of the Text, used for source-lineage context. Agreement among these rows is not treated as independent replication.

[7] Executed local record. EDKSG-VOY-V019-BOUNDARY-LINEAGE-01. The package contains the nine-relation source-lineage ledger, conditional-likelihood reproduction, conservative intersection sensitivity and exact hashes. It contains no high-resolution manuscript image and performs no source correction.

Collection: Voynich Research Library · Editorial framework: Wintour House.

EDKSG-VOY-V019 · Edition 1.0.0 · 5 September 2026 · Baseline EDKSG-VOY-B000. Claim-scoped boundary-lineage sensitivity audit. No image adjudication, complete-source verification, independent replication or semantic decipherment is claimed.