Publication note — 5 September 2026: Released with the publisher’s approval as exploratory research. Independent review remains pending. The research calculations and their stated limitations are unchanged.
Voynich | edkSG Research Volumes · EDKSG-VOY-V005 · Edition 1.0.0 · 5 September 2026, Singapore
Research status: completed exploratory analysis of inherited development excerpts. Not a full-corpus result, independent replication, translation or completed decipherment.
Where this volume fits
Volume 4 — The Geometry Check established how to preserve different text locations without pretending that every location belongs in the same experiment. This volume takes the next step: it constructs one explicit boundary comparison and asks whether the apparent clue survives a more demanding comparison of like with like.
Research record · Reviewed research status · Browse by subject
A clue can look impressive because it is common in the right neighbourhood. The harder question is whether it still distinguishes alternatives when the neighbourhood is held constant.
In our selected transcription excerpts, a preceding form ending in y is associated with the next eligible form beginning qo rather than o. The pooled table gives an odds ratio of approximately 5.03. That is an actual calculation from the retained sample, not a translation and not a reproduction of the older reported boundary figures. [2]
But the next check changes what we can reasonably conclude. When we group observations by the same page, recorded hand and current form after removing its optional leading q, only six groups containing 15 observations include both prefix outcomes and both predecessor-ending classes. Every one of those observations comes from f115r. Adding the same within-location ordinal leaves three groups containing six observations. [2]
The association is worth investigating. Its incremental value beyond context is not yet established. The result is neither “we have found the rule” nor “the clue is worthless”. We have identified the precise comparisons that are missing before the stronger claim can be tested convincingly.
1. One question, not another theory of the whole manuscript
The question in this volume is deliberately narrow: among eligible neighbouring transcription spans, does the preceding span’s final y help distinguish a current oR form from its qoR counterpart after accounting for recorded context?
Here, R is simply the remaining nonempty transcription string. It is not a decoded root. The Roman letters are transcription labels, not established sounds. A span is an analytical unit created under a stated separator policy; calling it a word would introduce an additional assumption.
The test also distinguishes two questions that are easy to confuse. We examine whether the previous span ends in y while the current span begins o or qo. We do not compare a current form with and without its own final y. The inherited TEXT.SIGMA_Y.001 question therefore remains a separate, uncompleted experiment in this record. [1, 2]
Nor are we testing line starts. Every admitted transition stays inside one supplied text-location record. Nothing is joined across two records, two paragraphs or two pages. The starting span of a location has no preceding span in this comparison and cannot contribute a line-opening observation.
This precision matters because a broad phrase such as “the y effect” can conceal several different computations. A result about the preceding ending is not automatically evidence about the current ending, a paragraph reset, a grammatical suffix or a reading instruction.
The purpose is to let one claim face a clear challenge. The broader language, cipher, notation and structured-generation possibilities remain open.
2. The source population stays visible
The input comprises the two local excerpts retained by the previous volumes: 191 text-location records across eight page units. They cover f1r, f1v, f2r, f2v, f115r, f67r1, f67v1 and f68r1. The samples were originally selected to develop and challenge extraction behaviour, not to estimate manuscript-wide linguistic properties. [1, 2]
These are inherited browser-transferred excerpts. Their local bytes are retained and identified, but they are not a newly acquired complete ZL transcription file. Fresh byte-level identity with the complete upstream source remains unverified in this run. Reusing them permits an explicit exploratory analysis; it does not permit an exact full-corpus replication claim.
The first filter retains only P-type locations with recovered marked-paragraph context. It admits 112 records, sets aside 76 non-P records, and retains but excludes three unmarked P records from this experiment. The distinction between preservation and analytical eligibility established in Volume 4 remains intact. [2]
The main analysis splits doubtful seams when forming spans, but admits a transition only across a confident dot separator. Drawing interruptions and doubtful separators are not bridges. Both participating spans must consist solely of ordinary lowercase transcription characters, without alternatives, ligatures, uncertain marks, extended codes, apostrophes or attached inline comments. An excluded span is not deleted to join its neighbours.
The current span must be o followed by a nonempty remainder, or qo followed by that same kind of nonempty remainder. Bare o and qo do not enter the contrast.
| Accounting stage | Count | Unit being counted |
|---|---|---|
| Preserved input | 191 | Text-location records |
| Eligible marked P input | 112 | Text-location records |
| Adjacent candidate pairs in those records | 887 | Span pairs under the split policy |
| Admitted oR/qoR comparisons | 201 | Eligible transitions |
These counts have different units. There are not 887 independent pages, nor 201 independent experiments. Some locations contribute several transitions, and the same span can participate in neighbouring candidate pairs. The ledger preserves that dependence rather than hiding it behind a large total. [2]
3. The pooled result is real within its scope
The 201 admitted transitions produce this table:
| Preceding span | Current qoR | Current oR | Total |
|---|---|---|---|
| Ends in y | 76 | 28 | 104 |
| Does not end in y | 34 | 63 | 97 |
| Total | 110 | 91 | 201 |
All four cells are calculated from the retained event ledger. Each event carries its page, locus, recorded hand, current matching key, ordinal position and the offsets of the participating source components. [2]
Within this selected sample, 76 of 104 transitions after y lead to qoR, approximately 73.1%. Among the other predecessor endings, 34 of 97 lead to qoR, approximately 35.1%.
The descriptive odds ratio is:
(76 × 63) ÷ (28 × 34) = 5.0294…
That is approximately five times the odds, not five times the probability. It is also not a probability that a linguistic theory is correct. It describes the relative composition of the two rows of this table.
We have not attached a population significance claim or a confidence interval. The sample is purposively selected, previously inspected and highly dependent. A familiar statistical label would not repair those conditions. It would be particularly misleading to present the number as a randomly sampled estimate of the entire manuscript.
The archived research contains older reported boundary odds ratios around 5.87 and 6.20. This is a newly defined calculation with explicit present rules. Similar numerical magnitude does not demonstrate that the earlier procedure has been recovered. The original event tables and analysis code behind those older figures remain unlocated in this run. [1]
The legitimate new statement is: this defined local comparison contains an association. Everything stronger needs another step.
4. Much of the table comes from one page
Of the 201 observations, 148 come from f115r. The page contributes 102 of the 110 qoR outcomes. Those are approximately 73.6% of the admitted observations and 92.7% of that outcome class. [2]
That concentration immediately limits how broadly the table can speak. The number of rows is not the number of independent production contexts. A densely represented page can dominate a pooled result even when the displayed list of source pages looks varied.
The inherited annotations distinguish two hands on f115r. Within their respective local subsets, the unadjusted odds ratios are 3.60 and 7.20. Those are additional descriptions of the selected table; they do not establish an effect independent of the exact current form or other contextual differences. [2]
Importantly, concentration alone does not prove that the whole association is an artefact. We reran the selection after removing f115r entirely. The remaining 53 observations produce an unadjusted odds ratio of approximately 2.21. A positive pooled association remains. [2]
What disappears outside f115r is more specific: there are no same-page, same-hand, same-current-form groups containing both prefix variants and both predecessor-ending classes under this policy.
This is the difference between measuring something in a heterogeneous remainder and possessing the comparison needed to distinguish two explanations. We can still count a pooled tendency. We cannot use that count as though the missing like-for-like contrasts had been observed.
The page-removal check therefore does not demonstrate “no effect”. It identifies where the available contextual comparison is concentrated, and where it is absent.
5. What matching actually asks
Suppose one current form is naturally common after y and another current form is naturally common after a different ending. A pooled table could show an association even without the same form changing its prefix choice in response to the predecessor.
To examine that possibility, we retain the exact current form except for an optional initial q. Thus oR and qoR share the analytical matching key oR. We group events by that key, their page and their recorded hand.
This is not a linguistic claim that q is a prefix in the historical language. It is an operational comparison of two strings. The proposed relationship is allowed to fail without requiring us to decide what either string means.
The split-policy data contain 156 such groups. Nine contain both current prefix outcomes, accounting for 22 observations. Only six also contain both predecessor-ending classes, accounting for 15 observations. These are the groups with the variation needed for the narrow within-group contrast being examined. [2]
| Comparison scope | Groups with both prefix outcomes and both predecessor classes | Observations in those groups |
|---|---|---|
| Same page, hand and current matching key | 6 | 15 |
| Also same within-location ordinal | 3 | 6 |
| Same page, hand and current key, after removing f115r | 0 | 0 |
The second row adds ordinal position within the supplied text location. It does not match a measured horizontal position on the manuscript. It also depends on the separator policy: joining a doubtful seam can change the ordinal of later spans.
The rows are diagnostics of comparison support, not estimates of a causal effect. We do not compute a headline “fully adjusted” ratio from the tiny remainder. We also do not pretend that a group lacking one of the required variations disproves the proposed relationship.
Other contextual differences remain unaddressed. We have not simultaneously matched preceding lexical identity, all aspects of paragraph position, physical spacing, every production variable or a securely established historical word class. More controls can reduce the available overlap still further.
The result is useful precisely because it shows the difference between an attractive pooled number and the much smaller set of observations capable of answering the stronger question.
6. Two cases prevent the simplest rule from becoming a translation
A universal rule would be much stronger than a probabilistic tendency. If the proposal were “a preceding y always selects qoR”, a single retained counterexample would matter.
Consider two source-addressed cases under the split policy:
| Source location | Recorded hand | Current ordinal | Previous span | Current span |
|---|---|---|---|---|
| f115r.1 | 2 | 10 | oechedy | oteedy |
| f115r.4 | 2 | 10 | sheod | qoteedy |
The matching key is the same, oteedy. The first predecessor ends in y, yet the current form is oR. The second predecessor does not end in y, yet the current form is qoR. We checked these short sequences against the currently readable ZL display as well as the local event record. That is a targeted textual cross-check, not a complete-file hash verification or inspection of the manuscript image. [2, 3]
These examples defeat that deterministic wording within the tested representation. They do not defeat every possible probabilistic or context-dependent relationship. They also do not justify selecting a new exception after every inconvenient case until the rule becomes impossible to challenge.
This is why we preserve the individual cases beside the aggregate. A large table can support a tendency while its own rows refute a universal law. Both statements can be true.
Nothing here authorises a translation of q as an action, grammatical operation, instruction or meaning-bearing word. We are comparing represented forms. The semantic question remains open after the statistical question has been made more precise.
7. Changing the doubtful-space policy does not supply an independent replication
The transcription format distinguishes confident gaps from doubtful gaps and drawing interruptions. We retain those source distinctions and generate three explicitly named analytical views. The format definitions are external conventions; the decision about which transitions enter this experiment is ours. [4]
In the split view, doubtful seams divide spans, but the transition under study must still cross a confident dot. In the joined view, doubtful seams join the material within a span. In the stricter view, any pair touching a doubtful seam is excluded. Drawing interruptions remain barriers in all three.
| Analytical policy | Admitted transitions | Pooled descriptive odds ratio | Context-overlap observations | Also same ordinal |
|---|---|---|---|---|
| Doubtful seams split | 201 | 5.03 | 15 | 6 |
| Doubtful seams joined | 204 | 5.27 | 14 | 2 |
| Pairs touching doubtful seams excluded | 183 | 4.68 | 12 | 4 |
The broad pooled association remains positive under these three treatments. That is a useful sensitivity observation. It is not three independent transcriptions, three fresh manuscript samples or three independent confirmations of a historical mechanism. The views share the same inherited source material. [2]
Their denominators also differ. A policy changes which spans exist, which pairs are eligible and where an ordinal falls. It is therefore incorrect to regard every change in the ratio as the isolated effect of punctuation while pretending the sample has remained identical.
The more restrictive comparison remains small under every policy. All context-overlap observations still lie on f115r. Removing that page leaves no such overlap in any of the three views. The apparent robustness of the pooled table has not supplied the missing distribution of contextual contrasts.
That is the central lesson of the sensitivity check: a clue can survive several reasonable preprocessing choices while still being inadequately tested against its strongest contextual alternative.
8. What the other research branches contribute
The wider library helps decide which challenges this clue must face. It does not add independent observations to the 201-event table.
The Class-vs-Identity branch asks what exact identity contributes after broader structure is known. We use that distinction here by comparing a pooled association with a stricter same-form comparison. We have not rerun that branch’s external class-induction study, and none of its reported numerical results enters this analysis. [5]
The Padua feature-ablation branch proposes removing a complete evidence family rather than one decorative feature. Here the practical analogue is removing the whole page supplying most of the observations. The resulting loss of matched support is an executed local result. It is not a geographical finding, and it does not complete the earlier localisation experiment. [6]
The Tangential and period-eye work separates compatibility from discrimination. A familiar-looking clue is useful only when competing explanations make different predictions about it. In this experiment, the relevant competition is not between named medieval languages. It is between an additional predecessor relationship and explanations based on the represented current forms and contexts. [7]
The Comparator branch keeps documented manuscript behaviour separate from Voynich identity. A known manuscript can suggest how to think about heterogeneous records without becoming evidence that Voynich shares its function. No comparison manuscript has supplied a decoded label or semantic answer in the present run. [8]
The object and custody branch keeps the representation’s limits visible. These computations concern transferred transcription strings and inherited annotations. They do not measure a physical gap, establish ink order, reconstruct a foldout’s handling or settle the original sequence of the codex. [9]
Finally, the numbered research volumes supply the reusable source and processing history. We are not beginning from an imaginary clean slate. We are using the preserved excerpts and their declared limitations to ask the next narrower question. [1]
This cross-branch review covered the principal research fronts and selected relevant articles, not a reacquisition of every source behind every published page. Its outcome is a better-designed comparison, not a claim that all earlier hypotheses have converged on one explanation.
9. The result is an evidence map, not a winning theory
We can now give each candidate statement a distinct position.
The selected local comparison has a pooled association: supported by the saved counts under the stated rules. The two-by-two table, policy variants and source-addressed events are available for inspection.
A preceding y invariably selects qoR: contradicted by retained cases under this representation. A tendency must not be rewritten as an exception-free instruction.
The predecessor ending adds a generalisable effect beyond current form and context: not established. The available within-context comparisons are too few and too concentrated to sustain that claim. No held-out prediction was performed.
The association is wholly explained by page composition: also not established. The pooled remainder after removing f115r remains positive, and the present work has not fitted and validated a complete contextual explanation.
The relationship identifies language, cipher, purpose or meaning: not tested. Several mechanisms could be compatible with the descriptive association. Compatibility has not selected a historical identity.
This is not indecision disguised as caution. It is a record of which questions the actual observations answer and which ones require a different sample or procedure.
The most important new constraint is practical: further effort should target the missing like-for-like comparisons, rather than merely accumulating more unbalanced instances of an already visible pooled pattern.
10. What was checked, and what was not
The new analyser passed 28 constructed software tests. They exercise separator barriers, uncertainty exclusion without bridging, prefix definitions, matching keys, metadata handling, sample boundaries, ordinal comparisons, source offsets and the two-by-two arithmetic. These are engineering checks, not 28 independent manuscript experiments. [2]
A second counting implementation reconstructed the split-policy admitted pairs directly from the preserved payload strings rather than using the first span classifier. It agreed on all 201 event identities, positions and participating strings. The source-component offsets were also checked for every admitted event in all three policy views. [2]
Both implementations share the inherited location and paragraph-context fields, and both were prepared within this investigation. Their agreement is an internal reproducibility check. It is not independent expert review or independent palaeographic assessment.
A short design record was written before the first calculation in this run. It named the question, candidate explanations, three spacing policies and forbidden overclaims. That record is not an external preregistration. The source excerpts and earlier questions were already known, and the overlap diagnostics are exploratory. The final implementation’s conservative treatment of attached inline comments is recorded explicitly rather than silently folded into an imagined preregistered method.
An early verification script used the wrong field name when requesting the preserved payload and stopped before comparison. The field-name error was corrected and the successful comparison rerun. That debugging event did not change the source, the eligibility rules or the numerical findings; it remains in the run history.
We have not validated every possible input syntax, independently checked the entire source transfer, or tested a held-out manuscript population. Those limits remain attached to the output even when the local tests pass.
11. The next collection should be designed for disagreement
The next stage needs a verified source edition and a population capable of distinguishing the surviving explanations. More rows alone are not the objective.
First, obtain and preserve the intended complete transcription bytes, or declare a different input edition explicitly. Reconcile the source identity and the parser’s accepted locations before describing the analysis as complete or as a reconstruction of the older experiment.
Second, construct a coverage map for the current-form families. Ask where both oR and qoR occur, which predecessor-ending classes are represented, and whether comparable cases extend across more than one physical or production group. The present study’s concentration on f115r makes that map particularly important.
Third, define the contextual comparison before inspecting the decisive outcome. Match only variables that the representation genuinely provides. A recorded hand is not a new expert identification; an ordinal is not a physical coordinate; a missing value is not permission to infer the preferred classification.
Fourth, distinguish discovery material from evaluation material. These 191 records have now been used repeatedly during development. They cannot become an untouched test set by receiving a new filename. Any later generalisation claim needs genuinely withheld material and an account of dependence between development and evaluation groups.
The next test should compare a model using the declared context with one adding the predecessor-ending feature, on the same eligible evaluation population. It should report both predictive changes and the population excluded by its rules. A failure to improve prediction would weaken that bounded proposal; inadequate overlap would identify another coverage problem. Neither outcome should be renamed as semantic discovery.
The visual experiment remains separate. There is no reason to select a picture after seeing a token result and then present that picture as independent confirmation. A future text–image connection must earn its own source alignment, controls and evaluation.
The checkpoint we can carry forward
Volume 5 moves beyond asking whether the representation survives extraction. It asks what a specific pattern is entitled to claim once a more demanding comparison is introduced.
The answer is clear enough to guide the next step. We have an exploratory pooled association, visible counterexamples to its simplest deterministic interpretation, and a precise shortage of matched comparisons outside one page. The stronger context-independent and semantic claims remain unestablished.
The useful clue is not the one with the most impressive isolated number. It is the one that can distinguish its nearest alternatives under a comparison we can actually make.
That is the result added to the dated research record.
Sources and research record
[1] Predecessor and inherited state. Vol. 4 — The Geometry Check, WordPress 154086, and its retained research package. The sequence is traced through the exact title and URL of Vol. 3 — The Extraction Gate, not the separate source-identity article bearing the same Volume 3 number. The current-token final-y question and older reported boundary statistics remain inherited, not completed by this study.
[2] Executed local analysis. EDKSG-VOY-V005-BOUNDARY-EXPLORATORY-01. The accompanying research package retains the two excerpt identities, design record, preserving-reader dependencies, analyser, source-addressed candidate and admitted-event ledgers, all three policy results, page-removal checks, test log and internal verification report. Counts are computed from the retained local representations. Full-corpus coverage and independent review are not claimed.
[3] Targeted source cross-check. Zandbergen–Landini ZL3b-n, version header 3b, 13 May 2025; browser display checked 5 September 2026. The short f115r.1 and f115r.4 examples above were checked at their source addresses. This check does not identify the complete currently served bytes or verify the manuscript images.
[4] Representation conventions. René Zandbergen, IVTFF specification, sections 6.7–6.9; relevant PDF pages inspected. The specification distinguishes separator and metadata conventions. The eligibility policies and comparisons in this volume are separately declared analytical choices.
[5] Structure and identity route. The Class-vs-Identity Problem. Used as a methodological continuation; its external numerical claims were not reproduced here.
[6] Ablation route. The Feature Ablation Test. Its localisation protocol is not reported as executed by this volume.
[7] Observer and comparison route. What Would Look Normal in 1400 but Strange to Us?. Historical compatibility is not treated as identification.
[8] Comparator boundary. Wellcome MS.683: The Working Receptarium. No historical details from that comparison are used as observations in the statistical sample.
[9] Object boundary. What Can Only the Object Tell Us?. The present study remains transcription-based and reports no material examination.
Collection: Voynich Research Library · Editorial framework: Wintour House.
EDKSG-VOY-V005 · Edition 1.0.0 · Prepared 5 September 2026 · Baseline EDKSG-VOY-B000. Original editions are retained; later corrections should identify the affected result, representation and edition. This research report is not an independent replication or decipherment.