Voynich | edkSG Research Volumes · EDKSG-VOY-V014 · Edition 1.0.0 · 5 September 2026, Singapore
Research status: executed fixed-model challenge on three additional Currier-A, recorded-hand-1 page samples. Two samples cannot supply the defined conditional comparison. This is not a certified historically unseen test, independent scientific review, full-corpus analysis or decipherment.
Three new page samples do not necessarily produce three new tests.
We selected f15r, f17v and f88r from recorded quires absent from the preceding event samples. All three carry Currier A and effective hand 1 in the inherited frame. The same models that had just scored three Currier-B samples positively were waiting for them, with their coefficients unchanged.
The acquired sources yield 77 eligible transitions. Yet only one same-form group containing five observations can score the defined conditional relationship. It lies on f88r. The other two pages cannot supply that comparison at all. [2]
The frozen common model improves the score of that five-observation group by approximately +0.378 against zero. That is a bounded positive result. It is not three more successful transfers, and the two unscored pages are not failures of the relationship.
A closer audit identifies the single unprefixed occurrence that makes the comparison possible. Without it, the group has no variation in the current prefix outcome. We retain that occurrence and its original source address; we do not remove it to change the result.
The new constraint is on our ability to test the clue: a page can contain substantial writing, and even a positive pooled association, without containing the like-for-like alternatives required by this experiment.
1. Challenge the successful model, not just its strongest examples
Volume 13 fixed two models before retrieving additional source objects. Its common coefficient improved the conditional score on all three target samples under all three declared gap treatments. Those samples were nevertheless selected, unequal and all assigned Currier B. The result justified a further challenge, not a universal conclusion. [1]
The present challenge changes the sampling population while preserving the learned coefficients. We use the identical frozen-model file, trained on the earlier 685-event collection. The 269 events added by Volume 13 are not added to the training set for this test.
The common coefficient remains approximately 1.544, with an odds-ratio equivalent of 4.683. The partial-sharing model’s fixed coefficient for hand 1 is approximately 0.907, equivalent to an odds ratio of 2.476. These are inherited model parameters. We do not re-estimate either value from the new pages. [2]
The target remains the predecessor-y relationship in represented oR and qoR forms. R is a nonempty remainder in a transcription string; an optional initial q is removed only to define an exact current-form family. None of that establishes a root, sound, grammatical prefix or historical meaning.
Changing the target to Currier A is not a claim that two languages have been identified. Here A is an inherited classification used to define the sampling contrast. The source headers and frame provide that label; this run does not independently rediscover its categories or identify what caused their differences. [3]
2. A selection rule recorded before acquisition
The sampling rule uses only the inherited 227-unit metadata frame and the exposure record available before these source files were fetched.
First, exclude every recorded quire already represented in the 685-event training collection, the Volume 13 targets and the earlier geometry-only page f67v1. Then, in the retained frame order, take the first Currier-A, hand-1, M1 page with at least fifteen P-type locations from each of the first two remaining qualifying quires. Finally, take the first qualifying M5 page from another remaining quire, with the same classification and minimum location count.
This deterministic rule selects f15r, f17v and f88r. It was written to the local design record before their target source objects were retrieved and before their transition tables were calculated. No substitute page was specified or introduced after the results became visible. [2]
| Page unit | Recorded group | Recorded quire | Inherited visual class | P / L locations |
|---|---|---|---|---|
| f15r | Q2-B2 | 2 | M1 | 15 / 0 |
| f17v | Q3-B1 | 3 | M1 | 23 / 0 |
| f88r | Q15-B2 | 15 | M5 | 16 / 15 |
The fifteen-location minimum is a practical attempt to avoid extremely short source samples. It does not promise fifteen usable transitions, or even one usable conditional comparison. That distinction becomes a central finding of the experiment.
The broader research history also remains unresolved. Being absent from the retained event sample does not establish that nobody previously examined a page. The design is locally pre-acquisition and purposive, not externally preregistered or certified historically blind.
3. Three exact derivative objects, not the complete parent file
The origin-host and raw-mirror downloads again failed at name resolution in the working environment. They contribute no new source bytes. We record an access failure, not a checksum disagreement or an excuse to call the full-transcription task complete.
The connected repository route returned three complete diplomatic objects. Each includes the page header, location number, location code, raw-IVTFF column and cleaned column. The local reconstructions were admitted only after their computed Git blob identities matched the identities returned by the connector. All three matched. Their SHA-256 identities are recorded separately. [2, 3]
Only the raw-IVTFF fields enter the analysis. The cleaned fields are retained locally for whole-object verification, not substituted for uncertain or interrupted source text. A precise copy of a cleaned column would still be a different analytical input.
The files contain 69 locations: 54 P-type records and fifteen labels. Their per-page counts and paragraph-start totals of one, one and three reconcile with the inherited frame. All 69 raw payloads reconstruct exactly from the preserving parser’s event representation. [2]
These checks identify the source representation being tested. They do not authenticate every glyph reading against a manuscript image. Three exact derivatives are not a fresh hash of their complete origin file, and multiple files from one transcription lineage are not independent readings of the manuscript.
4. The eligible transitions are a much smaller population
The extraction, geometry and boundary modules are unchanged from the preceding procedure. The main view splits doubtful seams into spans, but admits a pair only across a confident dot separator. Drawing interruptions remain barriers. Alternatives, ligatures, uncertainty and attached annotations exclude the affected plain-span pair. Excluded material is not deleted to connect its neighbours.
The current form must be o followed by a nonempty remainder or qo followed by a nonempty remainder. Bare o and qo do not enter the contrast. Every pair remains inside one supplied location record. Neither page boundaries nor separate lines are joined to create additional observations.
The 54 P-type records supply 309 candidate pairs. Of those, 77 are admitted, 181 lie outside the oR/qoR contrast, 37 are blocked by doubtful or drawing-related boundaries, and fourteen contain a nonplain or annotated span. These are mutually exclusive first-reason dispositions in the saved ledger. [2]
f15r contributes six admitted events, f17v contributes 36, and f88r contributes 35. The minimum source-location count has therefore not produced a balanced statistical sample.
In particular, all six admitted f15r events have a predecessor not ending in y. This does not mean the page contains no y characters or y-ending forms. It means none of those forms participates as the predecessor in the six events admitted under this experiment’s rules.
5. What the pooled tables say
| Page unit | y→qo | y→o | non-y→qo | non-y→o | Pooled OR |
|---|---|---|---|---|---|
| f15r | 0 | 0 | 3 | 3 | Undefined |
| f17v | 1 | 4 | 2 | 29 | 3.625 |
| f88r | 3 | 6 | 12 | 14 | 0.583 |
| Total | 4 | 10 | 17 | 46 | 1.082 |
The batch total is much closer to one than the corresponding total in Volume 13. But this is not an isolated test of Currier classification: the selected pages, recorded hands, current-form composition and other contextual features differ together. We cannot assign the change to A versus B from this design.
The f17v table is especially instructive. Its pooled ratio is positive, yet no exact current-form family in that page’s admitted sample supplies both prefix outcomes. The pooled association does not create the missing same-family comparison.
f88r shows the reverse-looking combination: its pooled ratio is below one, while the one available matched family scores positively under the frozen model. That is not a contradiction. The pooled table and the conditional score evaluate different questions over different usable subsets and weights.
Neither summary should cancel the other. The whole page’s descriptive tendency is not the same quantity as the arrangement within the one family permitting our fixed-model test.
6. Two pages are unscored, not unsuccessful
The conditional evaluation matches exact page, recorded hand and current-form family. Within each such stratum, it conditions on the observed number of qo outcomes. It then compares the observed allocation under the frozen coefficient with the allocation under coefficient zero. This removes a stratum intercept; it does not predict an unread token without knowing the stratum’s outcome total. [4]
A stratum whose current outcomes never vary cannot provide this comparison. If all observed members of a current-form family are oR, or all are qoR, the conditional allocation is fixed. Such a stratum does not supply evidence about the coefficient merely because it contains repeated text.
Neither f15r nor f17v has an outcome-varying exact-form stratum in the admitted sample. Their recorded status is NO_EVALUATION_INFORMATION, with no score value. They are not assigned a zero that could later be interpreted as a tie, a success or a measured absence of effect.
f88r supplies the single evaluable stratum, containing five observations. Both the common and the hand-specific models are compared on those exact same five observations.
| Page unit | Admitted events | Conditional observations | Common gain | Hand-specific gain |
|---|---|---|---|---|
| f15r | 6 | 0 | Not scored | Not scored |
| f17v | 36 | 0 | Not scored | Not scored |
| f88r | 35 | 5 | +0.378 | +0.272 |
The model has not passed two silent tests. It has encountered two missing comparisons. That result concerns what our selected evidence and matching rule can answer, not whether the manuscript possesses a particular kind of structure.
7. All the score comes from one family
The sole matched family is okol / qokol on f88r. These names are transcription strings, not translated terms.
| Source location | Previous span | Current span | Predecessor ends in y? |
|---|---|---|---|
| f88r.19 | qockhol | okol | No |
| f88r.27 | chol | qokol | No |
| f88r.27 | qokol | qokol | No |
| f88r.29 | chey | qokol | Yes |
| f88r.29 | chody | qokol | Yes |
Four current outcomes are qoR and one is oR. Two predecessors end in y, and both of those are paired with qoR. A positive coefficient therefore gives the observed arrangement more conditional probability than zero does.
This is a small pattern with a precise address. The two f88r.27 events also share a span, so even the five observations should not be described as five independent pieces of historical evidence.
The gain has a simple closed form. For coefficient β, it is log(5) − log(3 + 2 exp(−β)). At the fixed common coefficient it is approximately 0.377743. Even as a positive coefficient grows without limit, this particular gain cannot exceed log(5/3), approximately 0.511.
The bound is mathematical, not an inferential confidence limit. It shows how much information this one arrangement can contribute under the declared conditional score. It is a reason not to let a positive sign do more rhetorical work than the tiny comparison permits.
8. An image-checking priority, not a reason to discard an occurrence
After inspecting the primary results, we performed a separately labelled influence diagnostic: remove each of the five contributing occurrences in turn, without refitting either model, and ask whether the same conditional comparison remains available.
Removing the occurrence at f88r.19 leaves four current outcomes, all qoR. The family then has no outcome variation and the entire selected batch becomes unscorable for this conditional test. The other four single-occurrence removals leave a scored comparison. [2]
This does not show that f88r.19 is mistaken. No image has been examined in this volume, and no transcription has been corrected. Its leverage is a reason to verify it carefully, not a reason to delete it because the score depends on it.
A balanced image-level audit should inspect this unprefixed occurrence alongside the four prefixed ones and relevant excluded alternatives. It should ask whether the source readings, gaps and location alignments are supported, rather than asking an image reviewer to confirm the desired model.
The influence diagnostic was designed after seeing the result. Its timing is preserved. It does not replace the primary selection, change the primary table, or turn a targeted source-check priority into new palaeographic evidence.
9. The spacing alternatives do not create missing comparisons
The three target views were fixed before acquisition. The split view divides doubtful seams; the joined view combines them; the stricter view excludes candidate pairs touching such a seam. Drawing interruptions remain barriers throughout. The same frozen coefficients are used in every view.
| Target policy | Admitted events | Conditional observations | Common gain | Hand-specific gain |
|---|---|---|---|---|
| Split doubtful seams | 77 | 5 | +0.378 | +0.272 |
| Join doubtful seams | 77 | 5 | +0.378 | +0.272 |
| Exclude touching pairs | 71 | 4 | +0.219 | +0.161 |
Both unscored pages remain unscored under all three policies. The stricter view removes one occurrence from the sole informative family, leaving a smaller positive comparison on f88r.
The batch’s pooled ratio changes from approximately 1.082 to 0.788 under the stricter view. Its conditional score remains positive. Again, these are different summaries of changing eligible populations, not rival translations.
The common model scores better than the hand-specific model in every declared view here. That is a local ordering based on one sparse family. It does not settle the comparative value of hand-specific modelling elsewhere, and the repeated views do not multiply the number of historical observations.
10. Checks completed in this run
The original frozen-model file is byte-identical to Volume 13’s. The training collection still contains 685 observations; the previous 269 evaluation events are retained separately and never used for refitting. Selection is recomputed from the frame and must agree with the saved pre-acquisition rule.
The three source objects pass complete Git blob and local SHA-256 checks. All 69 payloads round-trip, and the primary event extraction preserves source offsets. A second raw-text path reconstructs all 77 admitted primary events, participating strings and offsets without invoking the primary event extractor.
A separate statsmodels likelihood calculation agrees across twelve available score comparisons, with maximum absolute difference approximately 4.5 × 10−16. Twelve additional page-model-policy checks confirm the unscored state, rather than attempting to fit or score empty comparison populations. These totals cover different jobs. [2, 4]
Thirty-four new unit and regression tests passed, alongside 28 inherited boundary tests. They include source tampering, frozen-model tampering, selection drift, duplicate rejection, invalid predictor values, original-location checks, exact sparse-case arithmetic, undefined-ratio handling, no-refitting execution and preservation of the unscored states.
These are internal computational checks. Both implementations share source material and were prepared in this investigation. They do not establish independent scientific review, image accuracy or historical independence.
11. What changes in the catalogue
The compatible primary archive grows from 954 to 1,031 distinct admitted transitions. Directly informative same-form groups increase from 48 to 49, and the observations in them increase from 242 to 247. Seventy-seven additional events have therefore added only one new directly informative group. [2]
Twenty-two page units now appear in the retained development history, including the earlier geometry-only page. Twenty-one contribute events. Their fifteen recorded bifolio groups cover seventy units in the 227-unit metadata frame; the remaining 157 have exposure not established by this audit.
Those 157 are not certified unseen. The new targets are now explicitly development-exposed, and their source-related neighbours require the same care before any future evaluation claim.
The cumulative counts describe the archive after this experiment. They are not a larger training set for today’s models. Nor do we add conditional gains from different historical fitting arrangements into a single apparent success score.
The stopping point is consequently traceable: source objects acquired, fixed models evaluated where possible, two missing-comparison states preserved, and a specific image-verification priority recorded.
12. The next question should address the missing comparison
The new batch does not justify a claim that Currier A lacks the predecessor relationship, that hand 1 behaves uniformly, or that a positive model fails on two of the three pages. Those two pages did not provide the defined test.
It also does not justify saying the model passed another three-page challenge. The positive evidence comes from five observations in one family on one page. Both overstatements would turn a coverage limit into a result.
The next useful work can proceed along two bounded routes. One is source-to-image verification of the influential family, including favourable and unfavourable observations. The other is a broader source census to locate genuine same-family contrasts in this under-supported population. The latter must remain discovery work until a later evaluation set and procedure are frozen.
Complete parent-source acquisition is still required for a complete transcription-wide opportunity map. The repeated network obstacle remains open; these three successful derivative-file admissions do not close it.
The clue has not disappeared. The opportunity to test it has become much narrower than the number of new pages suggested. Recognising that difference is the research advance in Volume 14.
Sources and reproducibility record
[1] Exact predecessor. Vol. 13 — The Frozen-Coefficient Check, WordPress post 154267, edition 1.0.0. Its frozen models, 685-event training index, 269-event evaluation index and inherited parser lineage are retained with their identities.
[2] Executed local record. EDKSG-VOY-V014-COUNTER-CONTEXT-01. The accompanying package contains the pre-acquisition selection and model bindings, source identities, non-plaintext event indices, scores, opportunity map, separately dated influence diagnostic, tests, numerical checks and publication record. No target refitting or image-level verification is asserted. Complete third-party diplomatic files are not redistributed; exact acquisition instructions are included.
[3] Source objects and metadata. Zandbergen–Landini-derived diplomatic files in cipher_benchmark: f15r, f17v and f88r, retrieved 5 September 2026. Git blob and local SHA-256 identities are retained. The metadata frame is the inherited 227-unit representation carried by the predecessor package, not a new codicological examination.
[4] Conditional-model reference. statsmodels: ConditionalLogit, consulted 5 September 2026. Used for the conditioning interpretation and numerical comparison, not as evidence for Voynich semantics.
[5] Evaluation boundary. scikit-learn: Cross-validation, evaluating estimator performance, consulted 5 September 2026. Its general separation of fitting, selection and evaluation informs the protocol; it does not certify manuscript-group independence or historical blindness.
Catalogue continuity: Vol. 3A — The Extraction Gate and Vol. 3B — The Source Identity Cross-Check retain their original URLs and historical identifiers. Earlier editions and the public hub are not rewritten by this publication.
Collection: Voynich Research Library · Editorial framework: Wintour House.
EDKSG-VOY-V014 · Edition 1.0.0 · 5 September 2026 · Baseline EDKSG-VOY-B000. Fixed-model counter-context extension with explicit missing-comparison states. No independent scientific review, certified historically unseen test, full-source verification or decipherment is claimed.