Voynich | edkSG Research Volumes · EDKSG-VOY-V010 · Edition 1.0.0 · 5 September 2026, Singapore
Research status: completed retrospective cross-page transfer diagnostic on the compatible preserving-method event collection. This is not a preregistered held-out validation, independent scientific review, full-transcription analysis or decipherment.
A pooled association can be large and still be a poor travelling rule.
That is the central result of this volume.
The compatible preserving-method record now contains 685 eligible transitions across 15 contributing page units. Pooled without exact current-form matching, the predecessor-y association gives an odds ratio of approximately 5.19. Restricting the descriptive comparison to exact page, recorded hand and current-form families leaves 33 informative strata containing 180 observations, with a Mantel–Haenszel summary of about 4.31. [2]
Those numbers preserve the positive structural clue. They do not answer whether the clue transfers.
We therefore performed a leave-one-page-out diagnostic. For each page with an evaluable conditional comparison, we fitted one common predecessor-y coefficient on the other pages, then froze that coefficient and asked whether it improved the conditional likelihood of the omitted page compared with a zero-effect model. The answer is mixed: 5 pages improve and 4 worsen. Summed across those folds, the fitted effect improves conditional log-likelihood by about 8.85, but much of that gain comes from f75r and f106r. [2]
The association has transferable signal in this retrospective collection, but it is not a uniform page-independent rule.
This is a narrower and more useful statement than either “the pattern disappears after context” or “we have found the rule”.
1. From association to transfer
Volumes 5 through 9 progressively tightened the question. The early pooled association was strong. Matching exact current-form families reduced the number of directly comparable observations. New pages then showed that the local association can be weak, strong, or even slightly reversed.
Once a relationship varies across pages, one more pooled table is not enough. A stronger test asks whether a coefficient estimated somewhere else helps on the page currently withheld from fitting.
The target remains represented strings of the form oR and qoR, where R is a nonempty remainder and the optional leading q is removed only to define an exact current-form matching key. This is not a linguistic root. The Roman letters are transcription labels. A preceding span ending in y is a represented structural condition, not an assigned meaning.
The model therefore has one deliberately small claim: within exact page, hand and current-form strata, does the binary predecessor-y feature help discriminate the current qo outcome?
Nothing in the transfer test interprets q as an action, prefix, sound or instruction. Nothing interprets y as a suffix, state marker or word ending in a historical language. The result is about a relation among encoded transcription features under one preserving representation.
The distinction matters because transfer can fail for several reasons: the proposed relationship may be wrong, the relationship may be context-dependent, the matching representation may be incomplete, or the available page may contain too little variation. A failed fold is evidence against a universal transfer claim, not a translation of why the fold failed.
2. The input is larger, but still historical development material
The input combines the 379 compatible split-policy events entering Volume 8 with the 306 additional events produced in Volume 9. Event identities are unique across the combination. The primary method is the preserving split policy retained after the representation correction in Volume 7. [1–3]
The resulting collection contains 685 events across fifteen page units that contribute at least one admitted event. It is uneven. Some pages contribute only a handful of events; f75r, f106r and f115r contribute much more.
Every page in this collection has already been exposed to the research programme. The leave-one-page-out procedure withholds one page from a particular model fit, but it does not make that page historically unseen. The model family, matching rule and question were developed after earlier inspection of related material.
For that reason, this volume uses the term retrospective transfer diagnostic, not held-out validation. The computation mimics one useful feature of external testing—fitting without the evaluation page—but does not recreate a truly untouched research history.
This limitation is not cosmetic. A genuinely prospective validation would freeze the source edition, parser, matching variables, coefficient form and evaluation pages before their target outcomes were inspected.
3. The pooled association remains strong
The complete 685-event table is:
| Preceding span | Current qoR | Current oR | Total |
|---|---|---|---|
| Ends in y | 231 | 108 | 339 |
| Does not end in y | 101 | 245 | 346 |
| Total | 332 | 353 | 685 |
The descriptive odds ratio is 5.1884.
This is a useful summary of the accumulated selected collection. It is not an estimate from a random sample of the manuscript and should not be interpreted as the probability that a general rule is true.
The exact-form comparison is narrower. Grouping by page, recorded hand and current-form family yields 33 strata where both predecessor classes and both current outcomes occur, containing 180 observations. The descriptive Mantel–Haenszel ratio across those strata is 4.3114. [2]
A conditional logistic calculation over outcome-varying exact-form strata gives a common fitted odds ratio of approximately 4.91. Its conventional model-based diagnostic p-value is very small, but we do not promote that number into a population claim. The sample is purposive and previously inspected, its observations are clustered, and its historical independence is not established. [2]
More important for this volume is what happens when the estimated relationship has to travel.
4. How the transfer diagnostic works
For each evaluable page, the page is removed from fitting. The remaining pages supply a conditional logistic estimate for one coefficient: predecessor-y.
The conditioning strata are exact page, recorded hand and current-form family. Within a stratum, the likelihood conditions on the number of qo outcomes, so a separate intercept for each current-form context does not need to be estimated explicitly.
The fitted coefficient is then frozen and carried to the omitted page. On that page, we compute the conditional likelihood of the observed qo allocation within each outcome-varying exact-form stratum. We compare it with the likelihood when the predecessor coefficient is exactly zero.
A positive difference means the coefficient learned elsewhere assigns more conditional probability to the observed held-out-page arrangement than the zero-effect model. A negative difference means it assigns less.
This is not ordinary accuracy. There is no threshold such as “predict qo when probability exceeds 0.5”. It is a comparison of conditional likelihoods under two structural models while holding the exact current-form strata fixed.
Pages without any outcome-varying exact-form strata cannot contribute to this diagnostic. They are not classified as failures. The necessary conditional comparison simply does not exist in the retained sample.
5. Five pages improve; four worsen
| Omitted page | Evaluation observations | OR learned from other pages | Δ conditional log-likelihood | Direction |
|---|---|---|---|---|
| f100r | 7 | 5.53 | -1.391 | Worsens |
| f100v | 5 | 4.84 | 0.382 | Improves |
| f101r | 12 | 5.43 | -0.418 | Worsens |
| f101v | 8 | 5.16 | -0.161 | Worsens |
| f102r1 | 11 | 4.70 | 1.001 | Improves |
| f106r | 62 | 5.38 | 3.600 | Improves |
| f115r | 22 | 5.53 | 0.059 | Improves |
| f26r | 9 | 5.24 | -0.535 | Worsens |
| f75r | 86 | 2.85 | 6.317 | Improves |
The mixed pattern is the headline. The trained positive coefficient improves the omitted-page conditional likelihood for f100v, f102r1, f106r, f115r and f75r. It worsens the fit for f100r, f101r, f101v and f26r.
The largest gain is f75r at approximately +6.32 conditional log-likelihood units. f106r contributes another +3.60. Their positive transfer is large enough that the total across the nine evaluable folds remains +8.85 despite the four negative folds.
This is not a count of five votes against four. The pages contain very different amounts of evaluable information. f75r contributes 86 observations to the conditional evaluation; f100v contributes five. Treating them as equal Bernoulli trials would discard the structure of the computation.
Nor does a positive fold prove that the same historical mechanism produced every contributing event. It says that one coefficient learned from the other current data is better than zero for that page under this conditional model.
6. The negative folds are part of the result
The negative folds matter because they directly constrain the simplest travelling-rule interpretation.
On f100r, the coefficient learned elsewhere has an odds-ratio equivalent of about 5.53, yet it worsens the conditional likelihood by about 1.39 units compared with zero. f26r also worsens by about 0.53. f101r and f101v worsen more modestly.
These pages do not establish an opposite universal rule. Their evaluable samples are small, and several contain very few informative same-form groups. The correct conclusion is not “the effect reverses on four pages”. It is that one positive coefficient does not improve every page.
This distinction protects both sides of the evidence. A strong pooled association should not erase counterperforming contexts. A few small negative contexts should not erase larger positive transfers elsewhere.
A future explanation must account for the unevenness rather than treating it as noise simply because a global average is convenient.
7. The overall transfer gain is concentrated
The total transfer improvement across the nine evaluable pages is positive, but it is heavily influenced by the two largest positive pages. f75r and f106r together contribute about 9.92 conditional log-likelihood units—more than the net +8.85 after the negative and smaller positive folds are included.
That concentration is not automatically a defect. If the underlying relationship is genuinely stronger in certain contexts, a correct model should discover that. The problem arises only if we describe the common coefficient as though it transfers uniformly.
Removing each informative page from the descriptive Mantel–Haenszel summary provides another view of concentration:
| Page removed | Remaining informative strata | Remaining observations | Descriptive MH OR |
|---|---|---|---|
| f100r | 32 | 177 | 5.06 |
| f100v | 32 | 175 | 4.22 |
| f101r | 31 | 168 | 4.55 |
| f101v | 32 | 175 | 4.42 |
| f102r1 | 31 | 174 | 4.09 |
| f106r | 21 | 120 | 4.35 |
| f115r | 27 | 165 | 5.12 |
| f26r | 32 | 174 | 4.48 |
| f75r | 26 | 112 | 2.78 |
The summary stays above one no matter which single informative page is removed. The largest reduction occurs when f75r is removed, leaving an odds ratio of about 2.78. Removing f106r leaves about 4.35.
This makes two facts coexist. The accumulated association is not solely the product of one page, but one page can materially change the size of the summary.
The appropriate next question is therefore not whether the effect exists in one averaged sense. It is which recorded or still-unrecorded contextual variables explain the change in strength.
8. A common effect is an approximation, not the result
Across the 33 informative exact-form strata, a conventional Breslow–Day equal-odds diagnostic rejects perfect homogeneity at approximately p = 0.0072. Because the strata are sparse, selected and previously inspected, we treat this as a warning about the adequacy of a single common-effect description—not a manuscript-wide inferential theorem. [2]
The leave-one-page-out result gives a more tangible version of the same warning. A coefficient fitted on the other pages can help a large page substantially, do almost nothing on another, and hurt a third.
Thus the phrase “the predecessor-y effect” is becoming too coarse. The project has evidence for a conditional association family whose observed strength changes across represented contexts.
That does not yet identify the conditioning variable. Recorded page, hand, broad visual class, Currier population, paragraph position, exact preceding form, copying state and other local structure remain possible contributors. Some are inherited scholarly classifications; others would require new measurements.
We should resist a new shortcut: noticing that f75r and f106r differ from f26r does not make their visual classes or scribal hands the cause. A metadata label becomes explanatory only when a test separates it from the variables travelling with it.
9. Why this matters for the research pathway
A pooled success can be a false pass for a stronger claim.
The pooled 685-event table clearly contains a positive association. The exact-form summary also remains positive. If the research question were only “is there some association in this accumulated representation?”, those results would be enough to answer yes within scope.
But the stronger claim is about transfer: can a rule learned in one part of the represented manuscript improve our account of another part without being retuned?
The answer is not a simple yes. It improves five evaluable pages and worsens four. The net improvement is positive, but concentrated. The simplest page-independent travelling rule therefore receives partial support and a clear limitation at the same time.
This is precisely why the numbered record is useful. Volume 5 could preserve the original clue. Volume 7 could correct a representation mismatch. Volume 8 could measure comparison opportunity. Volume 9 could add new contexts. Volume 10 can now test whether the relationship transfers rather than protecting the original impression.
Progress here is not the uninterrupted strengthening of one hypothesis. It is the sharpening of what that hypothesis is allowed to claim.
10. The transfer result still does not supply meaning
A coefficient that transfers between pages would still be a structural relationship. It would not, by itself, tell us why q is present, why y occurs, or whether the units correspond to language.
The observed dependency could arise in a linguistic system, a copying or encoding procedure, a state-dependent generative process, a transcription artefact, or a combination. The current analysis does not discriminate all of those mechanism families.
The negative folds make semantic overreach especially risky. If a proposed translation says that preceding y instructs q to appear, the pages where that simple transferred relation performs poorly would require principled explanation. Adding ad hoc semantic exceptions after each failure would make the proposal difficult to falsify.
A better route is to predict the structural exceptions first. If another measured variable explains why the relationship is strong in one page and weak in another, that variable becomes a candidate part of the generating mechanism. Meaning can then be investigated on a more constrained system.
We therefore retain q and y as transcription-level features in this article. No semantic label is promoted.
11. The next experiment: predict the strength, not merely the sign
The next useful research question is now more specific. Instead of repeatedly asking whether the association is positive, ask which available context predicts its strength and transfer performance.
A bounded next model should begin with variables already represented independently of the target result: recorded hand, Currier population, broad page class, position information and exact previous-form families where coverage permits. The model must record how much direct overlap remains after each addition.
The objective is not to find the combination producing the largest retrospective coefficient. It is to define a small set of competing context models, freeze them, and evaluate their transfer on genuinely reserved material after the exposure history is reviewed.
A useful model may conclude that the predecessor-y feature adds value only under a subset of contexts. Another may show that a simpler representation of current-form composition accounts for most of the apparent transfer. A third may remain inconclusive because the necessary contrasts are absent.
All three outcomes would improve the research state because they narrow the mechanism without assigning meaning prematurely.
Complete source reconciliation remains a parallel requirement. A model whose statistics become more sophisticated than its input provenance has solved the wrong layer of the problem.
12. What was checked in this volume
The two parent event indices are identified by SHA-256 in the research record. The combined index contains exactly 685 unique event identities. The 33 informative exact-form strata contain exactly 180 observations.
The common conditional model is fitted only on outcome-varying exact-form strata. Each leave-one-page-out fold refits without the evaluation page, freezes the coefficient, then evaluates the page through a direct conditional-likelihood calculation rather than refitting to that page.
The calculation enumerates the possible allocations of the fixed number of qo outcomes within each evaluation stratum and compares the observed predecessor-y alignment with those alternatives. The zero-effect model uses coefficient zero under the same conditional population.
Internal assertions verify the event count, stratum count, informative-observation count, number of evaluable pages and five-improve/four-worsen result. The complete result ledgers and code are retained with the research package. [2]
These are computational verification steps, not independent scientific replication. The same research programme developed the representation, selected the prior samples and designed the retrospective transfer test.
The rule travels—but not everywhere
At this checkpoint, the most concise summary is:
The predecessor-y / qo association is not merely a pooled artefact of one page, and a coefficient learned elsewhere can improve the conditional fit of several omitted pages. But the improvement is uneven and fails on four of nine evaluable pages. A single page-independent rule is therefore too strong for the evidence we currently have.
That conclusion preserves the clue without converting it into a decoding key.
The pathway has moved from detecting an association to testing its portability. The next stage should explain the portability differences.
Sources and reproducibility record
[1] Immediate predecessor. Vol. 9 — The New-Context Check, EDKSG-VOY-V009. Its 306 primary split-policy events and new-context source record enter this diagnostic unchanged.
[2] Executed local research record. EDKSG-VOY-V010-TRANSFER-01. The accompanying package contains input identities, the 685-event combination, exact-form stratum tables, common conditional-model output, leave-one-page-out conditional-likelihood ledger, page-removal summaries and executable calculations. The statistical diagnostics are retrospective and not promoted to population-level proof.
[3] Earlier compatible evidence. Vol. 8 — The Comparison-Opportunity Map, and through it the corrected preserving-method records from Volumes 5 and 7. Volume 6’s cleaned-column event set is not added as an independent source.
Catalogue: Voynich Research Library · Editorial framework: Wintour House.
EDKSG-VOY-V010 · Edition 1.0.0 · 5 September 2026 · Baseline EDKSG-VOY-B000. Prior editions remain preserved. No independent review, certified unseen evaluation, semantic decoding or manuscript-wide effect is claimed.