Voynich | edkSG Research Volumes · EDKSG-VOY-V007 · Edition 1.0.0 · 5 September 2026, Singapore
Research status: completed retrospective representation-consistency audit. This volume corrects how the preceding comparison should be interpreted; it does not announce a decipherment or a new held-out validation.
The previous calculation can be reproduced. The previous comparison needs a qualification.
Those are different findings, and both matter.
Volume 6 reported 157 transitions, a pooled odds ratio of about 2.69, and a same-folio, same-current-form stratified estimate of about 1.63. We reconstructed all 157 saved events from the repository’s cleaned text column and recovered those numerical results. The table was not an arithmetic invention. [1, 3]
But the processing was not the preserving procedure used in Volume 5. Along with introducing different pages, Volume 6 used text in which drawing interruptions and ligature brackets had already been removed and alternative readings had been resolved. Its dot-only split also left comma-containing spans to be excluded by an alphabetic-character check. The population and the representation policy changed together. [2–4]
The correction is therefore specific: Volume 6 is a reproducible exploratory result in its declared cleaned representation, but it is not an unchanged-method validation of Volume 5. The five pages were outside the earlier sample; they were not an untouched evaluation set tested under a fully frozen, identical processing procedure.
Applying the unchanged Volume 5 procedure to the raw-IVTFF column of those same five page units produces 178 eligible transitions, a pooled descriptive odds ratio of 2.63, and a stratified descriptive estimate of 1.95. Seven informative groups contain 31 observations. Neither the older nor the new calculation establishes a general decoding rule. [3]
This volume adds the missing comparison: the same pages, an explicitly identified source column, and recoverable processing rules.
1. Five source objects are now locally identifiable
The audited page units are f100r, f100v, f101r, f101v and f102r1. They are the same five used in Volume 6, not a newly discovered or independently selected sample. The source repository provides one diplomatic text file for each. Each row contains a location number, a location code, a raw-IVTFF field and a cleaned first-alternative field. [4]
This time, we reconstructed the complete five diplomatic files from the decoded content returned by the GitHub connector. We admitted each local file only after calculating its Git blob identity and matching it against the identity returned by the connector. All five matched. Separate SHA-256 values identify the retained local files in this audit’s manifest. [3]
That is a useful advance over an unchecked manual transfer. A mistaken character, omitted label or altered line ending would ordinarily change the computed content identity and stop the admission check. The match applies to these five repository objects and their complete local contents.
It does not establish the accuracy of the transcription against the manuscript. It also does not close the outstanding identity check for the complete ZL3b source file. We still have not acquired and hashed the full origin-host byte sequence in this run. Direct origin retrieval failed; the successful route here was the small, individually addressable repository files.
The distinction must remain explicit: exact copies of five derivatives are not an exact copy of their complete parent. Repository preservation, transcription accuracy and historical interpretation are separate questions.
The five files contain 104 location records. Fifty-two are P-type records used by the present comparison; the other 52 are retained but excluded. All 52 P records have marked-paragraph context under the inherited reader. The page headers supply hand 1 and Currier A throughout this selected set. These are source assignments, not new palaeographic findings. [3, 4]
2. Reproducing the old result before changing anything
An audit should not begin by replacing a result it has not understood. We therefore first recreated the simplified Volume 6 procedure from the cleaned column.
Within each selected P record, the procedure splits at periods. It accepts neighbouring spans only when both pass the alphabetic-character check. It then keeps current forms beginning o with a nonempty remainder, or qo with a nonempty remainder. The predictor is whether the previous accepted span ends in y. No adjacent records are joined.
The resulting events match the saved Volume 6 ledger in their page, P-line index, ordinal, previous string, current string, predecessor class, outcome and current-form matching key. This is stronger than obtaining the same total: different event sets can produce identical totals, so we checked the actual recorded events. [3]
| Preceding span | Current qoR | Current oR | Total |
|---|---|---|---|
| Ends in y | 24 | 22 | 46 |
| Does not end in y | 32 | 79 | 111 |
| Total | 56 | 101 | 157 |
The pooled ratio is (24 × 79) ÷ (22 × 32), or 2.6932. The descriptive stratified ratio reconstructed from the seven saved informative groups is 1.6257. We retain these as results of that procedure, rather than retroactively claiming the older calculation used today’s rules.
One addressing issue also becomes clearer. Volume 6’s line index counted P records after other record types were excluded. On f100r, P-line index 1 corresponds to source locus 12, not source locus 1. The new crosswalk retains both values. An analysis index remains useful, but it must not masquerade as an original source address. [3, 4]
3. What the cleaned column had already decided
The repository documents its benchmark transformations. Its cleaned representation selects the first alternative reading and removes markup. Those can be legitimate choices for a particular benchmark, provided the question and limitations are stated. They are not automatically the right choices for every boundary experiment. [6]
The raw fields in our 52 selected P records contain nine drawing-interruption markers across six records, eleven alternative-reading constructs, seventeen ligature constructs and 48 commas. These counts concern the selected raw representation, not the whole manuscript. [3]
Two different kinds of information can disappear during cleaning. One is uncertainty: after a bracketed alternative has been resolved, a later alphabetic-character test cannot detect that the source originally offered alternatives. The other is segmentation: after a drawing interruption has been deleted, the resulting letters can appear to form one uninterrupted span.
The IVTFF specification distinguishes periods, commas and drawing interruptions. It also distinguishes uncertainty and ligature notation. These are transcription conventions. They do not supply historical meanings, but an experiment concerned with boundaries needs to know which conventions it has preserved or changed. [5]
Volume 6 did disclose its use of the cleaned column. The problem is not that the source was secretly different. The problem is that the surrounding interpretation could make the move from Volume 5 sound like a test of the same procedure on new pages. It was not. The same statistical target does not make two input-construction procedures identical.
This correction is directed at our comparison, not at the external repository. The repository preserved the raw column alongside the cleaned one, making the present audit possible.
4. A source-addressed example of a changed current form
At f102r1.9, the raw field includes the sequence orsheoldy.qokeol<->oteeody. The cleaned field joins the material around the drawing interruption, yielding orsheoldy.qokeoloteeody. [4]
The simplified cleaned-column procedure counts a transition from orsheoldy to qokeoloteeody. The preserving procedure instead keeps the transition from orsheoldy to qokeol and does not bridge the drawing interruption.
Both current forms begin qo. The difference can therefore leave the same broad outcome class while changing the exact current-form family used for matching. That is why an apparently harmless cleaning change can affect a contextual analysis even when the pooled headline moves little.
We have not discovered that qokeol has a meaning, or that the drawing interruption is a linguistic sentence boundary. We have shown that two recorded representations supply different current-form identities to the calculation.
The same principle applies to a ligature-derived span. After its braces disappear, the cleaned letters may pass an ordinary alphabetic check. Under the Volume 5 policy, that source span remains ineligible for the plain-span contrast. Exclusion here records a conservative analytical choice; it is not a judgement that the underlying writing is meaningless.
The essential research object is therefore not merely a pair of strings. It is the pair together with its source location, source representation and admission rule.
5. The preserving procedure was reused, not retuned
For the matched rerun, we reused the three relevant Volume 5 modules without changing their bytes: the extraction reader, geometry-preserving layer and boundary analyser. Their SHA-256 identities were compared with the inherited package. All three were unchanged. [3]
A small adapter constructs an IVTFF-shaped input from the diplomatic raw fields, restoring each supplied page header, original location number and location code. This wrapper is clearly labelled as a derivative. It is not presented as a newly downloaded ZL3b file.
The inherited reader preserves all 104 payloads. Concatenating each record’s parsed events reproduces its raw field exactly. The boundary analyser then uses the same marked-P eligibility and ordinary-span checks used in Volume 5.
Its primary view splits doubtful seams into spans but admits transitions only across confident dot separators. It does not bridge drawing interruptions. Alternatives, ligatures, uncertainty residue and attached inline comments remain reasons to exclude the affected pair from this plain-span comparison. Excluded material is not deleted to connect its neighbours.
That procedure is not guaranteed to be the best historical representation. Its immediate advantage is comparability: we know which rule set produced the new sample and can distinguish a change of pages from a change of algorithm.
The raw-to-wrapper adapter is new and is included in the audit. Reusing the analysis modules does not mean every step of the pipeline was old. It means the new adapter supplies the fields expected by an unchanged, identifiable analysis, with its own transfer and reconstruction checks.
6. The preserving rerun admits 178 transitions
Under the split policy, the 52 marked P records supply 645 adjacent candidate pairs. The analyser admits 178, excludes 338 as outside the oR/qoR contrast, excludes 72 because a participating span is not plain or is annotated, and excludes 57 because the intervening boundary is doubtful or drawing-related. These are mutually exclusive first-reason dispositions in the saved ledger. [3]
| Preceding span | Current qoR | Current oR | Total |
|---|---|---|---|
| Ends in y | 25 | 26 | 51 |
| Does not end in y | 34 | 93 | 127 |
| Total | 59 | 119 | 178 |
The pooled descriptive odds ratio is (25 × 93) ÷ (26 × 34), or 2.6301.
The number of admitted events is larger than before, even though the preserving procedure is stricter about several kinds of annotation. That is not a contradiction. The older dot-only procedure rejected comma-containing spans wholesale. Splitting a doubtful seam can reveal a neighbouring plain span whose transition across the next confident dot is eligible under the preserving rule.
We compared the two event collections as multisets keyed by original locus and the participating strings. They share 150 pair occurrences. Seven occur only in the cleaned calculation; 28 occur only in the preserving split calculation. Their totals reconcile: 150 + 7 = 157 and 150 + 28 = 178. This multiset comparison preserves duplicate counts but is not a claim that cleaned and raw character offsets are interchangeable. [3]
Five of the seven cleaned-only events involve ligature-derived strings that the plain-span rule does not admit. Two involve strings joined across deleted drawing interruptions. The 28 preserving-only events include restored post-comma comparisons and correctly separated current forms.
These differences explain why a similar pooled ratio does not demonstrate that the same observations were tested.
7. The same-form comparison remains small
Next we group events by page, recorded hand and current form after removing an optional leading q. This is a transcription-string match, not a claim that q is a historical grammatical prefix.
The split-policy rerun contains 121 groups. Twelve have both current outcomes, oR and qoR. Seven also have both predecessor-ending classes. Those seven informative groups contain 31 observations, compared with 29 in the simplified cleaned calculation. Adding the same within-location ordinal leaves only one informative group with two observations. [3]
The descriptive Mantel–Haenszel ratio across the seven informative groups is 1.9530. Its calculation combines the within-group cross-products, scaled by each group’s size. We report it as a description of this selected collection, not as a causal estimate or a validated prediction score.
Its difference from the old 1.6257 is a representation-sensitivity result. It does not mean the historical effect has suddenly become stronger. We changed which represented events and matching families entered the summary.
Nor does the smaller stratified ratio, relative to the pooled ratio, prove that composition caused the pooled association. The summaries apply different grouping and weighting, and the usable contextual comparison is sparse. A decline in an odds ratio is not, by itself, a demonstration of a causal explanation.
We do not carry Volume 6’s p-value into the corrected event set. It belongs to the old table and its assumptions. This audit reports descriptive estimates and comparison support; it does not produce a new population-level significance claim.
The research conclusion remains bounded: the selected data contain a pooled association, but a generalisable predecessor effect beyond the relevant context has not been established.
8. Three preserved-space policies, three declared samples
The inherited analyser supports three spacing policies. We ran all three without choosing a winner after inspecting their ratios. Each preserves the source distinction; each declares a different analytical treatment.
| Policy | Eligible events | Pooled OR | Stratified descriptive OR |
|---|---|---|---|
| Volume 6 cleaned-column reconstruction | 157 | 2.69 | 1.63 |
| Preserving: split doubtful seams | 178 | 2.63 | 1.95 |
| Preserving: join doubtful seams | 176 | 2.68 | 1.83 |
| Preserving: exclude pairs touching doubtful seams | 152 | 2.68 | 2.45 |
The preserving split and joined views each have seven informative context groups, with 31 and 30 observations respectively. The exclude-touching view has six informative groups with 27 observations. They are not identical populations wearing different labels. [3]
All four pooled ratios are positive and fairly close in this selected sample. That makes the broad descriptive tendency worth retaining. The contextual summaries vary more, while the number of informative observations remains small.
The exclude-touching ratio of 2.45 is not preferred merely because it is larger. A researcher who chooses a policy because it increases the desired effect is selecting on the result. Conversely, choosing the smallest ratio as an automatic correction would be equally unjustified.
The table is a record of sensitivity, not a competition in which the most convenient number wins. The appropriate policy depends on the question, the representation and a procedure fixed before a genuinely evaluative result is examined.
These are also not four independent confirmations. They reuse five page units from one transcription lineage. More views help expose assumptions; they do not manufacture new historical witnesses.
9. What changes in the research record
Three conclusions are carried forward, and three interpretations are restricted.
Carried forward: Volume 6’s 157 events and numerical summaries reproduce under its documented simplified procedure. The five source files are now locally content-bound to their returned repository identities. A positive pooled association also appears when the Volume 5 preserving procedure is restored on those pages. [3]
Restricted: the transition from Volume 5 to Volume 6 cannot isolate the effect of introducing new pages, because the processing changed too. The term held-out should not be read as a fully frozen validation experiment. The lower stratified estimate is not proof that composition explains the association, or that the additional predecessor effect is absent.
This is a correction to comparability and interpretation, not an accusation that every old number was false. A reproducible calculation can answer a narrower question than the surrounding account suggests.
We also avoid the opposite overreaction. Recovering a source distinction does not automatically reveal the correct historical unit. The preserving procedure keeps uncertainty visible so competing analytical views remain possible; it does not settle their historical interpretation.
Earlier public editions are left unchanged in this action. This successor identifies the affected volume, procedure and claim. Future use of Volume 6 should carry this qualification rather than repeating the older inference without its processing context.
The unresolved current-token final-y experiment remains separate. Neither the previous cleaned analysis nor this rerun translates q or y, identifies a language, completes blind visual-component coding, or establishes a text–image semantic relationship.
10. Verification that was actually performed
The audit includes five complete-file content-identity matches, reconstruction of all 157 saved historical events, unchanged-code checks on three inherited modules, raw-payload round trips for all 104 locations and source-component offset checks for every admitted event in each preserving view. These are different checks serving different purposes. [3]
Eighteen new software tests passed. They cover source admission, original-locus crosswalking, old-table reproduction, drawing barriers, uncertainty and ligature exclusion, the three doubtful-seam policies and source-offset integrity. The 28 inherited boundary tests were also rerun successfully.
The tests verify specified software behaviour. They are not independent manuscript experiments, an expert review panel or a probability that the analysis is free of all errors. The two implementations share source material and were prepared within the same investigation.
The design note for this audit was written before its new calculations, but after the earlier result and its limitations were known. It is a retrospective audit plan, not external preregistration. The five pages are now part of the development history and must not be recycled as an untouched evaluation set.
The source-to-image comparison remains undone here. Repository content identity makes the computational handoff more precise; it does not make the underlying glyph readings certain.
11. The next experiment has to freeze the whole procedure
A future evaluation must bind more than the target question. It needs the source edition, source column, parser version, spacing policy, uncertainty treatment, eligibility population, matching variables and outcome calculation. A change to any of these may change what the test measures.
The immediate evidence priority remains complete source acquisition and reconciliation. The five verified diplomatic files demonstrate a workable small-file route, but they do not substitute for the full parent source or a manuscript-wide coverage map.
Once the intended corpus is available, the next useful output is an opportunity map: where do the same current-form families occur under both predecessor classes, and how is that overlap distributed across pages and production contexts? The number of comparisons that can discriminate alternatives is more important than the raw size of a token collection.
Development and evaluation material must then be separated without ignoring their physical and transcriptional relationships. All pages already inspected in these volumes belong in the exposure record. A newly named file is not newly unseen evidence.
The eventual predictive comparison should ask whether adding the predecessor-ending feature improves a context-only model on the same frozen evaluation population. It should report exclusions and remain open to a null, positive or inconclusive result. None of those outcomes should be converted directly into semantics.
The present checkpoint has made that future result more interpretable. It has not moved the finish line closer by declaring an intermediate calculation to be a solution.
A corrected pathway is stronger than an uninterrupted story
The temptation in a numbered research series is to make each volume sound more decisive than the last. This audit does something different: it checks whether the route between two results was as continuous as it appeared.
The old calculation survives reproduction. The implied same-method comparison does not survive without qualification. A matched rerun now supplies the missing reference, while leaving source accuracy, generalisation and meaning open.
Before asking whether a clue survived a new sample, establish that the procedure did not quietly become a different test.
That is the research advance recorded in Volume 7.
Sources and reproducibility record
[1] Audited predecessor. Vol. 6 — The Held-Out Confounding Check, edition 1.0.0, WordPress post 154105; its 157-event package was retained without modification.
[2] Preserving-method predecessor. Vol. 5 — The Matched-Context Check, edition 1.0.0, WordPress post 154103. Its three analysis modules were reused byte-for-byte. The earlier geometry and extraction work remains linked through that edition.
[3] This executed audit. EDKSG-VOY-V007-REPRESENTATION-01. The accompanying package contains the input manifest, reproducible acquisition instructions, analysis code, historical-result reconciliation, corrected event ledgers, sensitivity summaries, test logs and catalogue entry. No independent scientific review is claimed.
[4] Primary repository derivatives. The cipher_benchmark diplomatic files for f100r, f100v, f101r, f101v and f102r1. Retrieved through the connector on 5 September 2026. Full Git blob and local SHA-256 identities are recorded in the manifest. Attribution remains with the Zandbergen–Landini transcription and the repository’s documented transformation.
[5] Format authority. René Zandbergen, IVTFF specification, especially the uncertainty and inline-marker definitions and section 6.7. Relevant PDF pages were inspected. Representation conventions are not treated as decoded meanings.
[6] Transformation documentation. Voynich ZL3b Source Notes. The benchmark’s documented choices are distinguished from the preserving analytical choices used here.
Collection: Voynich Research Library · Editorial framework: Wintour House.
EDKSG-VOY-V007 · Edition 1.0.0 · 5 September 2026 · Baseline EDKSG-VOY-B000. Retrospective consistency audit and successor qualification of Volume 6. Prior editions remain preserved. No new held-out validation, complete-source verification, independent review or decipherment is asserted.