Voynich | edkSG Research Volumes · EDKSG-VOY-V009 · Edition 1.0.0 · 5 September 2026, Singapore
Research status: executed same-method exploratory extension on three additional page units. Source objects and calculation inputs are identified. This is not a certified held-out validation, independent scientific review, full-corpus result or decipherment.
The next three pages did not tell the same story.
Using one unchanged procedure, the pooled predecessor-y association is weakly reversed on f26r, strongly positive on f75r, and positive but smaller on f106r. The descriptive odds ratios are 0.70, 20.94 and 3.13, respectively. The first page contributes only fifteen eligible transitions, so its apparent reversal is especially fragile. [3]
The two larger pages retain positive descriptive associations after matching the current-form family within the page. The new data therefore do not support dismissing the clue merely because the earlier five-page comparison became weaker after matching. They also do not support announcing one uniform manuscript-wide rule.
The defensible advance is a more specific structural question: under which represented contexts does this relationship appear, and what else varies with those contexts?
We now have 306 additional eligible transitions, including 134 observations in twenty groups that support the defined same-form comparison. These observations come from three recorded bifolio groups not represented in the preceding event sample. That widens the comparison opportunity. It does not convert three selected pages into three independent historical experiments. [3]
1. The question has not changed with the new result
The target remains the same narrow relationship examined in Volumes 5 through 8. Among eligible neighbouring transcription spans, is a preceding span ending in y associated with a current form beginning qo rather than o?
The comparison uses the notation oR and qoR, where R is a nonempty remainder. Removing the optional leading q gives a common current-form matching key. This is an operation on represented strings. It does not establish a linguistic root, a grammatical prefix, a pronunciation or the meaning of either sign.
The current form’s own final y is a different question. Nothing in this volume completes the separately recorded current-final-y experiment. Nor are we examining paragraph or line starts: every admitted pair stays within one supplied text-location record.
Volume 7 established why keeping the method fixed matters. Its audit reproduced the preceding cleaned-column calculation, but showed that the earlier move to different pages had also changed the construction of the input. This new extension therefore reuses the preserving reader, geometry layer and boundary analyser byte-for-byte. [2]
The three new pages were selected and recorded before their target files were fetched and before their transition outcomes were calculated in this run. That gives the selection a useful local audit trail. It is not external preregistration, and it does not erase possible exposure during the wider research programme.
The primary question is allowed to produce an inconvenient answer. A page with the opposite descriptive direction stays in the report. A sparse comparison is not repaired by silently substituting a better-looking page.
2. Why these three page units?
Volume 8 identified thirteen directly used page units in six recorded bifolio groups. Its strongest direct comparisons were concentrated in only three of those groups. Returning repeatedly to the same densely represented page would add rows without adequately widening the contextual challenge. [1]
We therefore selected f26r, f75r and f106r using the inherited metadata frame. Each belongs to a recorded group outside both the preceding direct sample and its same-group neighbours. The selection intentionally varies the broad visual class and includes two recorded hands while remaining within the source’s Currier B classification.
| Page unit | Recorded group | Recorded hand | Inherited visual class | P-type records |
|---|---|---|---|---|
| f26r | Q4-B2 | 2 | M1 | 10 |
| f75r | Q13-B1 | 2 | M3 | 53 |
| f106r | Q20-B4 | 3 | M6 | 47 |
| Total | Three additional groups | Two assignments | Three inherited classes | 110 |
The source frame describes M1 as whole-vegetal drawing with associated text, M3 as human figures with fluid-, tube- or basin-like forms, and M6 as marginal-marker or text-dominant entries or panels. Those descriptions help explain the sampling contrast. They do not identify the pages’ semantic functions.
The grouping is also inherited. We have not newly inspected the binding or established that these three recorded groups are statistically independent. Two of the pages share a hand assignment, all three share Currier B, and all belong to one transcription lineage of one manuscript.
The design is consequently a purposive new-context challenge, not a balanced experiment comparing genres, scribes or Currier populations. Its small size is part of the interpretation, not something to conceal behind the number of token pairs.
3. What was actually acquired
The complete origin-host transcription remains unavailable to this working environment: the direct retrieval again failed at name resolution. We did not turn that access failure into a claim that the source changed or that its checksum disagreed with the archive.
Instead, the connected repository route returned three complete diplomatic files. Each contains the supplied header, location number, location type, raw-IVTFF field and cleaned field. We reconstructed the local objects and required their computed Git blob identities to match the identities returned by the connector. All three matched. Separate SHA-256 values identify the retained local bytes. [3, 5]
The cleaned fields were reconstructed only as part of the whole-file identity check. They are not the analytical input. The calculation reads the raw-IVTFF fields and preserves their distinctions before applying the unchanged eligibility rules.
That separation matters because the repository’s cleaned column can join text around a removed drawing interruption. The raw field remains available precisely so another research procedure can make a different, explicit choice. We use that retained field rather than trying to reverse cleaning after the information has disappeared.
The source objects contain 110 location records. Their page totals and paragraph-start counts agree with the inherited frame: ten records and two paragraph starts for f26r, 53 and four for f75r, and 47 and fifteen for f106r. All 110 records have marked P-type context under the inherited reader. [3]
These matches verify the computational handoff within the inspected representations. They do not independently verify the source readings against the manuscript image. Three complete derivative objects are also not the complete parent ZL3b file; the full-source identity task remains open.
4. The rules applied before any ratio was calculated
The primary view splits doubtful seams into spans but admits a pair only across a confident dot separator. Drawing interruptions remain barriers. Alternatives, ligatures, uncertain marks, extended codes and attached comments make the affected plain-span comparison ineligible. Excluded material is not deleted in order to join its neighbours.
The source format distinguishes confident gaps, doubtful gaps, drawing interruptions and paragraph markers. Those conventions describe the representation; our admission rule describes the experiment. Treating the two as separate avoids converting a transcriber’s notation directly into a historical linguistic claim. [6]
The current span must begin o with a nonempty remainder or qo with a nonempty remainder. Bare o and qo do not enter the comparison. A previous span is classified only by whether its represented final character is y.
The method therefore does not read through drawings, cross between records, infer a missing word or choose an alternative reading because it strengthens an effect. It also does not pretend that its admitted spans are known linguistic words.
The three inherited analysis modules match their previous SHA-256 identities. A new adapter supplies the additional diplomatic raw fields to those modules. Its source transfer, header mapping and payload preservation receive separate checks; unchanged analysis code is not a reason to leave the adapter unexamined.
The resulting population has 830 candidate pairs. The ledger admits 306, sets aside 382 outside the oR/qoR contrast, 121 at doubtful or drawing-related boundaries, and 21 involving nonplain or annotated spans. These are mutually exclusive first-reason dispositions, and they sum to the candidate total. [3]
5. The new pooled table
The 306 admitted events produce the following descriptive comparison:
| Preceding span | Current qoR | Current oR | Total |
|---|---|---|---|
| Ends in y | 130 | 54 | 184 |
| Does not end in y | 33 | 89 | 122 |
| Total | 163 | 143 | 306 |
The pooled odds ratio is (130 × 89) ÷ (54 × 33) = 6.4927…. This is an odds ratio, not a sixfold probability and not a probability that the proposed historical explanation is correct. [3]
It is also not a manuscript-wide estimate. The pages were selected to challenge contextual coverage, not sampled randomly in proportions representing the whole book. A pooled number can change when the mixture of pages changes, even with an unchanged parser.
No population p-value or confidence interval is attached to this table. The observations are clustered, some spans participate in neighbouring pairs, and the research population is deliberately selected. A familiar inferential label would not establish the missing sampling assumptions.
The pooled result supplies a reason to investigate. Its interpretation depends on the page-level breakdown and on the comparison that remains when current-form families are held constant.
6. Keeping the inconvenient page in view
| Page unit | y→qo | y→o | non-y→qo | non-y→o | Pooled OR |
|---|---|---|---|---|---|
| f26r | 7 | 5 | 2 | 1 | 0.70 |
| f75r | 81 | 21 | 7 | 38 | 20.94 |
| f106r | 42 | 28 | 24 | 50 | 3.13 |
The first row is important because it prevents the two stronger pages from becoming a universal story. On f26r, the direction is slightly below one, but the total is just fifteen. A single changed classification could materially alter such a small table. We do not use it to declare a stable negative rule.
The f75r row contains 147 events and the largest descriptive ratio. It is neither a vote from 147 independent pages nor evidence that every page with the same illustration class will behave similarly. Its repeated forms and local organisation deserve attention.
The f106r row contains 144 events and a positive ratio of 3.125. It supplies a second positive page-level result in a different recorded group and hand. That broadens the local evidence without making the pages independent or proving the hand caused the difference.
The contrast between f26r and f75r is especially useful: both carry hand 2 and Currier B in the inherited assignments, yet their selected tables differ greatly. The broad labels alone do not make their observed samples interchangeable. They also do not isolate the cause of the difference, because page length, form composition, layout and other variables have not been experimentally separated.
This is the appropriate level of the new question. We should investigate where the relationship changes, not announce that one metadata label has explained it.
7. What survives same-form matching?
For the contextual comparison, events are grouped by page, recorded hand and the exact current-form matching key. A group is called informative here only when both predecessor-ending classes and both current prefix outcomes occur. That does not require every cell to be nonzero and does not imply that the group contains semantic information.
The new sample contains 137 such groups in total. Twenty meet the informative-group criterion, containing 134 observations. Adding the same within-location ordinal leaves seventeen groups containing fifty observations. The ordinal is a position in the chosen span representation, not a measured physical coordinate. [3]
| Page unit | Informative groups | Observations in them | Descriptive stratified OR |
|---|---|---|---|
| f26r | 1 | 6 | 0.00* |
| f75r | 7 | 68 | 28.69 |
| f106r | 12 | 60 | 4.24 |
| New sample combined | 20 | 134 | 7.02 |
The stratified description uses the Mantel–Haenszel ratio, calculated as the sum of a×d/n across the informative groups divided by the sum of b×c/n. It is a weighted summary of these observed tables. We are not asserting a common causal effect or fitting a validated predictive model.
On the two larger pages, the positive descriptive association remains after this limited matching. That is genuinely different from the much weaker contextual summary of the earlier five-page extension. The new data constrain any claim that matching the current form necessarily removes the relationship everywhere.
They do not show that context has been fully controlled. The comparison does not simultaneously match previous-token identity, all line and paragraph conditions, physical spacing, copying history or a securely identified linguistic class. Unmeasured or unmodelled dependencies can remain.
The important change is evidential, not rhetorical: the archive now contains substantially more direct within-page, same-form contrasts, and those contrasts do not all point toward the same simple page-independent account.
8. Remove the strongest page
A natural concern is that f75r carries the new result. We therefore calculated the predeclared leave-one-page-out descriptions without changing the source or eligibility rules.
Removing f75r leaves 159 events from f26r and f106r. Their pooled ratio is 2.91; the descriptive stratified ratio is 3.85, based on thirteen informative groups and 66 observations. The positive combined description does not vanish with the strongest page. [3]
That is a useful robustness check, but it does not establish two independently corroborating positive contexts: the remaining positive evidence is largely supplied by f106r, while f26r remains small and differently directed.
Conversely, removing f106r leaves a much larger pooled ratio because the strongest page now dominates the remainder. Removing f26r also increases the combined ratio. Those changes are reasons to report the page mixture, not reasons to choose the mixture producing the largest result.
Leave-one-out analysis is a diagnosis of concentration. It does not create additional experiments. Every subset comes from the same already inspected three-page collection.
9. Could exact repetition be doing the work?
The source on f75r includes immediately repeated forms. Since a form can itself end in y and begin qo, an exact repeat can contribute directly to the positive cell of the table. This is an obvious alternative worth examining, rather than leaving it implicit.
After inspecting the primary result, we ran a separately labelled sensitivity check removing only events in which the previous and current represented strings are exactly identical. We did not remove near-repeats, invent a similarity threshold or change the primary analysis.
This removes six events: five on f75r and one on f106r. The remaining 300 events have a pooled ratio of 6.17 and a descriptive stratified ratio of 6.69. On f75r alone, the pooled ratio remains approximately 19.65. [3]
The narrow conclusion is that these six exact adjacent repeats are not sufficient to account for the whole observed association. A broader copy-and-modify mechanism, other local dependencies or unmeasured state could still contribute. This check does not refute an entire mechanism family.
Its timing is explicit: it was added after the primary result was seen. It is an exploratory diagnostic, not a preregistered success or a reason to replace the original result with a preferred subset.
This distinction also preserves the older research branch on repeated and modified forms. A simple explanation should face a test of its stated prediction, but failing one deliberately narrow version does not eliminate every more complex version.
10. The spacing checks tell a consistent but bounded story
The same inherited analyser was run under all three recorded treatments of doubtful seams. The split view forms separate spans; the join view combines the material around a doubtful seam; the exclude-touching view omits a candidate pair touching one. All continue to block transitions across drawing interruptions.
| Policy | Eligible new events | Pooled OR | Stratified descriptive OR | Informative observations |
|---|---|---|---|---|
| Split doubtful seams | 306 | 6.49 | 7.02 | 134 |
| Join doubtful seams | 308 | 6.42 | 6.67 | 129 |
| Exclude touching pairs | 268 | 6.49 | 5.47 | 117 |
The positive combined description survives these three choices, and the basic page-level contrast remains: f26r is small and below one, while f75r and f106r are positive. This makes the result less dependent on one of the tested doubtful-space treatments.
It is not three independent transcriptions or three independent tests of the historical mechanism. The views share source material and differ in their eligible observations and matching opportunities. The 306, 308 and 268 totals must never be added as though they were separate discoveries.
Nor do these checks establish robustness to every plausible representation. We have not tested another transcription alphabet, another scholarly transcription, image-based gap measurements or alternative glyph grouping. The inference stops at the choices actually examined.
11. What the comparison map gains
Volume 8’s primary preserving index contained 379 admitted events. The three additional page units do not overlap its sampled pages, and their event identities do not duplicate its entries. Combining the compatible split-policy indices produces 685 events.
The same-form comparison inventory increases from thirteen informative groups with 46 observations to 33 groups with 180 observations. Those informative observations now occupy six recorded bifolio groups, rather than three. These counts include contrasts in either direction; a group does not become informative only when it supports the preferred sign. [3]
This is the direct response to Volume 8’s finding. The work has widened the opportunity to compare, rather than simply multiplying observations in an existing group.
The exposure ledger also changes. Sixteen page units are now directly represented in the retained analyses, occupying nine recorded groups. Their group closure contains 44 page units, including 28 additional related units. The other 183 units in the archived frame remain exposure-unassessed, not certified unseen.
The new pages must now be treated as development history. They cannot later serve as an untouched validation set merely because today’s analysis is called exploratory. Their related pages require the same explicit group-aware review introduced in Volume 8.
No combined manuscript-wide effect estimate is inferred from the 685-event count. It remains an uneven, purposively accumulated collection. The increased coverage is real; representativeness has not been established.
12. Checks that were actually completed
All three reconstructed source files match their returned Git blob identities. The adapter preserves all 110 raw payloads. Reconstructing each payload from its parsed events gives an exact local match, and every admitted event’s source components were checked against their saved positions under all three policies.
A second implementation reads the raw text fields directly for the primary split policy, without calling the inherited scanner, span constructor or event extractor. It recovered the same 306 event identities, participating strings and source offsets.
The first version of that checker included the paragraph-end marker in the last span’s endpoint. It already agreed on the number of events, but failed the exact-offset comparison. The checker was repaired to track the end of the text itself; the primary parser, source bytes, eligibility rules and numerical results were not altered. A regression test now covers that case.
Twenty-two new source and regression checks passed, and all 28 inherited boundary tests were rerun successfully. The checks exercise source admission, page selection, payload reconstruction, frame reconciliation, group separation, accounting, uncertainty handling and the second counting path. [3]
These are implementation checks, not fifty independent manuscript experiments. Both implementations were prepared in this investigation and share the same source material. We have not obtained independent scientific review, verified the source against all relevant images or measured a predictive accuracy on unseen material.
The short research plan predates the new data fetches in this run. The added exact-repeat diagnostic does not. Keeping both timestamps and purposes visible is more useful than describing the whole process as though every decision had been fixed in advance.
13. What we can now say—and what still cannot be said
We can say that the observed relationship is not confined to the previously dominant f115r sample. Two additional page units show positive pooled and same-form descriptive associations under an unchanged preserving procedure.
We can say that the three new page samples differ markedly in their descriptive tables. The small f26r result must remain visible. We cannot yet say that those differences establish distinct historical mechanisms, because sampling variation, representation and other contextual factors have not been separated.
We can say that same-form matching does not automatically remove the association in these new contexts. We cannot say that it proves an independent causal action of the preceding y or a generalisable predictive improvement. Those are different claims requiring additional design and evidence.
We can say that six exact adjacent repeats do not fully account for the local pattern. We cannot say that copying, modification or state-dependent generation has been eliminated as an explanation.
We cannot translate the relationship. No meaning for q or y follows from an odds ratio. A structured writing process, a transcription convention, a linguistic dependence, a coding procedure or a combination could remain compatible with portions of the result.
The new evidence therefore corrects two temptations at once: the temptation to turn an attractive regularity into a decoding instruction, and the temptation to dismiss that regularity everywhere after one smaller contextual comparison weakens.
The working target is now a bounded context-sensitive structural relationship, with its historical explanation still open.
14. The next useful question
The full-source acquisition and identity check remains a prerequisite for claiming a complete corpus-wide opportunity map. Three successful derivative-file acquisitions do not close that task, and this volume does not report otherwise.
For the next bounded experiment, the new page contrast gives us a sharper design problem. We need to distinguish what varies with page composition, represented position, previous-form identity and production context, without introducing so many controls that the direct comparison disappears again.
A prospective model comparison should keep the source edition, raw-field treatment, eligibility rules and evaluation population fixed. A context-only model and a model adding the predecessor-ending feature must be judged on the same eligible observations. The pages already examined here belong in development, not in the final evaluation.
The resulting outcome may strengthen the structural claim, reveal a narrower condition, or show that an apparent predictive advantage does not transfer. The present results do not dictate which outcome the next test must produce.
The visual-component and current-final-y lanes also remain separate. We should not attach a suggestive illustration to the new table and count the resemblance as independent confirmation. Any text–image connection needs its own source alignment, blinded or otherwise clearly bounded assessment, and comparison against alternatives.
Volume 9 adds new evidence rather than a new metaphor. The clue now has a wider, uneven footprint. Understanding that footprint is a more useful next step than giving the clue a meaning it has not earned.
Sources and reproducibility record
[1] Exact predecessor. Vol. 8 — The Comparison-Opportunity Map, WordPress 154120. Its retained event indices, frame and exposure ledger supply the starting coverage, not new independent observations.
[2] Preserving method and correction. Vol. 7 — The Representation-Consistency Check, WordPress 154116, reusing the method recorded in Vol. 5 — The Matched-Context Check. The three inherited module identities are recorded in this run’s plan and were unchanged.
[3] Executed research record. EDKSG-VOY-V009-NEW-CONTEXT-01. The accompanying package contains the source manifest, acquisition instructions, fixed selection record, unchanged inherited code, new adapter and checker, hashed event indices, stratum summaries, spacing and page-removal results, post-result exact-repeat diagnostic, test logs and coverage update. Whole third-party transcription files are not redistributed in the package; source acquisition is separately required for a complete rerun.
[4] Inherited metadata. The 227-unit frame retained by Vol. 8, derived from the frozen matrix documented in Vol. 2 — The Input Audit. Page classes, hands and bifolio groups remain attributed assignments, not findings of a new image or object examination.
[5] Source objects. Zandbergen–Landini-derived diplomatic records in the cipher_benchmark repository: f26r, f75r and f106r, retrieved 5 September 2026. Exact Git blob identities and local SHA-256 values are in the manifest. The raw-IVTFF columns, not the cleaned columns, supply this analysis.
[6] Format reference. René Zandbergen, IVTFF specification, document issue 2.0.2, especially section 6.7 and inline-marker definitions. The relevant page image was inspected. This is a representation reference, not evidence for the manuscript’s meaning.
Catalogue: Voynich Research Library · Editorial framework: Wintour House.
EDKSG-VOY-V009 · Edition 1.0.0 · 5 September 2026 · Baseline EDKSG-VOY-B000. Prior editions are retained. The two works carrying the Volume 3 label remain disambiguated by exact title and URL in the earlier catalogue. No canonical model activation, independent scientific review, certified held-out validation or decipherment is claimed.