Voynich | edkSG Research Volumes · EDKSG-VOY-V017 · Edition 1.0.0 · 5 September 2026, Singapore
Research status: completed retrospective correction-scope audit of two compatible frozen-model evaluation batches. No new manuscript observations, model fitting, image-reading verdict or independent scientific review are claimed.
One disputed stroke can overturn a local result without overturning every result beside it.
The opposite mistake is equally serious: a positive total can conceal a local reversal and tempt us to stop checking the stroke.
This audit separates those two possibilities. Under the retained l-ending reading at f88r.19, the common frozen model’s local conditional gain is approximately +0.378. Under the documented y-ending alternative, it is −0.906. The local sign changes, exactly as Volume 15 reported. [1, 2]
When that same source scenario is evaluated together with Volume 13’s compatible three-page batch, the combined gain changes from +6.071 to +4.788. The combined sign does not change. Both calculations use the same frozen coefficients and the same 78 conditional evaluation observations under the primary policy. [3]
Both findings must survive the correction: the f88r result is reading-sensitive; the combined six-page index remains positive under this particular alternative. Neither finding chooses the glyph.
We have not acquired the image evidence needed to decide between l and y. We have instead completed an audit of the correction’s numerical scope, so that uncertainty in one observation is neither hidden nor allowed to invalidate unrelated results without examination.
1. Three claims that must not share one verdict
The first claim concerns the manuscript itself: what terminal form appears immediately before the relevant unprefixed okol? That requires source evidence at the correct location. It cannot be settled by which candidate produces a more attractive statistical score.
The second claim concerns a local analytical arrangement: does the fixed positive predecessor-y coefficient improve the conditional score of the five-case okol/qokol family on f88r? That answer depends on the represented terminal. Under the two attested readings, the local answers differ.
The third claim concerns an aggregate: does the same coefficient improve the sum of conditional scores across two compatible evaluation batches? That answer can remain positive even when one contributing family becomes negative.
A research record becomes misleading when these claims are compressed into “the result passed” or “the result failed”. The claims have different objects, evidence requirements and populations.
Volume 16 narrowed the source-location question. There is a label okol at f88r.15 and a running-text occurrence at f88r.19; only the latter belongs to the scored comparison. It also distinguished historical Currier hand numbering from the modern hand field. Those are useful completed alignment checks, not a completed terminal-stroke reading. [1]
The present audit does not repeat them as new discoveries. It follows their consequence: once the affected observation is identified, every downstream claim should state whether it actually depends on that observation and how.
2. The two batches are compatible in a specific way
Volume 13 evaluated f34r, f77r and f105r after fixing two models on the earlier 685-event training collection. Its primary split-policy batch contains 269 admitted transitions. Volume 14 subsequently evaluated f15r, f17v and f88r using the identical frozen-model file, adding 77 transitions. [2, 4]
Volume 15 preserved the second batch under three source treatments: the retained l reading, the documented y alternative, and a conservative analytical exclusion of the two directly affected pairs. We use those retained event indices rather than creating an additional transcription reading here.
The model files in the two parent packages are byte-identical. Their SHA-256 is recorded in this audit. The evaluation batches have disjoint page identities and recorded bifolio groups, and neither contains a recorded group used in the 685-event training index. These are checks of the retained metadata, not proof that the manuscript material is historically independent. [3]
The common coefficient remains 1.5440411135795047, with an odds-ratio equivalent of approximately 4.683. The partial-sharing model retains its previously fitted hand-specific coefficients. No part of the newly combined evaluation index is used to update either model.
The current-form matching key is also unchanged: the exact represented form after removing an optional initial q. That key is stored as a hash for equality comparisons. It identifies a transcription-string family, not a decoded root or a claim that the form has the same meaning everywhere.
This is a retrospective aggregation audit. Both component results and the disputed reading were known before its design was recorded. It is not a new prospective transfer test, and the act of combining the files does not make previously exposed material unseen.
3. Why these scores can be added—and other historical scores cannot
The score groups events by exact page, recorded hand and current-form family. For each stratum, it compares the conditional log-likelihood under a fixed coefficient with the conditional log-likelihood at coefficient zero.
The two batches do not share any such stratum. Since the coefficient used for each hand is fixed, regrouping their event-level union gives the same objective as summing the contributions of the two disjoint sets of strata. The calculation does not need a new fit.
Combined fixed-model score
= sum of stratum contributions from batch one
+ sum of stratum contributions from batch twoThis is an algebraic property of the declared objective. It is not an assumption that the source observations are independent historical witnesses. Neighbouring pairs can share spans, pages can share production history, and all of these records descend from one transcription lineage. Here the sum is a model diagnostic; we do not interpret its exponential as independently accumulated evidential odds.
The qualification matters because this series also contains leave-one-page-out and leave-one-group-out calculations. Their coefficients change between folds. Appending those totals to the present fixed-model score would combine different evaluation procedures, not enlarge one coherent test.
Likewise, adding the l scenario and the y scenario would count alternative representations of the same observations twice. Adding split-gap and joined-gap views would do the same. We recompute each scenario separately and retain its identity.
Earlier cautions against mixing evaluation totals therefore remain in force. This audit establishes a narrow exception through compatibility checks; it does not grant permission to total every favourable number in the research library.
4. Six pages, 346 events, 78 conditional observations
The primary retained-reading union contains 346 admitted transitions across six page units. Of those transitions, 78 lie in strata with both current outcomes. Sixty-seven observations, distributed across sixteen strata, also provide the predecessor variation relevant to the coefficient. [3]
The two denominators are not interchangeable. An outcome-varying stratum with a constant predecessor feature contributes no change between the fixed coefficient and zero. It can enter the likelihood accounting while supplying no direct information about that feature.
| Page unit | Admitted events | Conditional evaluation observations | Common score, retained reading |
|---|---|---|---|
| f34r | 29 | 2 | +0.500 |
| f77r | 107 | 34 | +2.545 |
| f105r | 133 | 37 | +2.649 |
| f15r | 6 | 0 | Not scored |
| f17v | 36 | 0 | Not scored |
| f88r | 35 | 5 | +0.378 |
| Total | 346 | 78 | +6.071 |
The 346-event union is an evaluation subset, not a replacement for the historical 1,031-event primary archive. No new manuscript observation is introduced, and no model is refitted to that archive.
Nor is this a six-out-of-six success. Only four pages have a scored comparison under the retained reading. f34r contributes one two-observation comparison, while the two larger positive contributions come from f77r and f105r.
Retaining these unequal contributions is essential to understanding why a local reversal need not reverse the sum.
5. The correction changes the local sign, not the combined sign
| Treatment at f88r.19 | f88r common score | Combined conditional observations | Combined common score |
|---|---|---|---|
| Retained l ending | +0.378 | 78 | +6.071 |
| Documented y alternative | −0.906 | 78 | +4.788 |
| Exclude both affected pairs | Not scored | 73 | +5.693 |
The unaffected three-page batch contributes +5.6932776974 in both scored reading scenarios. The local l-to-y change subtracts approximately 1.2833332869 from the combined score. That accounts exactly for the change between the first two totals, within floating-point precision.
This explains the positive aggregate without explaining away the negative local result. The two larger unchanged page contributions exceed the local loss. No mysterious compensation, revised coefficient or hidden extra observation is required.
Under the alternative reading, the page-level status is three positive scored pages, one negative scored page and two unscored pages. That is the correct descriptive inventory. It is not four successful confirmations merely because the total remains above zero.
Conversely, the local negative result does not establish that the frozen model fails on the other three pages. Their event inputs and their coefficients are unchanged, and their contributions reproduce unchanged.
The result therefore bounds a correction’s reach: it qualifies the specific local claim and reduces this aggregate score, but does not reverse the sign of this particular combined diagnostic.
6. A missing score cannot be disguised as a favourable zero
The conservative treatment excludes both event pairs whose representation directly depends on the disputed ending. One pair changes its current-form family when the preceding source token changes; the next pair changes its predecessor-y feature. The audit checks both effects rather than modifying only the convenient predictor.
Removing those two pairs leaves 344 admitted events in the combined split-policy index. But the sole unprefixed occurrence in the relevant f88r family is now absent, so that family no longer has the current-outcome variation required for scoring.
The local result is consequently NO_EVALUATION_INFORMATION, with a null score. The combined score remains available because the other batch still supplies 73 conditional observations. It equals that batch’s +5.693 contribution. [3]
There is a useful distinction between an algebraic contribution and a reported experimental result. A stratum with no possible outcome variation contributes no model-versus-zero difference to a sum. Reporting its page as a measured zero-score tie, however, would suggest that the necessary test had been performed. It has not.
The software retains both concepts explicitly: a zero formal contribution where the objective supplies none, and a missing local score where the comparison is unavailable.
Exclusion is not a permanent correction to the source. It is a sensitivity treatment for unresolved uncertainty. If suitable evidence later supports one reading, the corresponding source scenario can be reconsidered without pretending the interim exclusion was a palaeographic verdict.
7. Alternative gap treatments preserve the same scope distinction
The inherited gap policies are evaluated separately. The split view divides doubtful seams, the joined view combines them, and the stricter view excludes a pair touching one. All preserve drawing interruptions as barriers, and all use the same frozen coefficients.
| Gap policy | Combined common score with l | Combined common score with y | Combined score with affected pairs excluded |
|---|---|---|---|
| Split | +6.071 | +4.788 | +5.693 |
| Join | +5.214 | +3.931 | +4.837 |
| Exclude pairs touching doubtful seams | +5.953 | +4.689 | +5.734 |
The common model’s combined score stays positive in all nine defined scenarios. The local f88r score still changes from positive under l to negative under y in every gap policy. Under quarantine it remains unscored.
For the two attested reading scenarios across these policies, the smallest combined common score is approximately +3.931 and the largest +6.071. This is a range of explicitly calculated scenarios. It is not a confidence interval, a probability of a positive historical effect or an exhaustive bound over every possible transcription error.
We have varied one documented ending and three already-declared gap treatments. We have not varied all ambiguous glyphs, all segmentation systems or every influential observation in the two batches. A more extensive uncertainty analysis could have different results.
The scope statement is therefore precise: this identified uncertainty does not reverse this combined score under these tested views. The word “this” carries necessary scientific work.
8. Neither model may use the aggregate to select the source reading
The partial-sharing model is retained as a secondary comparison, with its frozen hand deviations unchanged. Its combined primary score is approximately +6.145 under l and +5.409 under y. It too remains positive after the local alternative reverses its f88r contribution. [3]
The source issue affects the two models by different amounts because their hand-1 coefficients differ. That can change their apparent relative performance without supplying evidence that either model identifies the true character.
It would be circular to prefer l because it increases the common model’s total, or prefer y because it changes the relative advantage of the hand-specific model. The target reading must be established through source evidence, not selected by a downstream model whose output depends on that reading.
Likewise, this retrospective sensitivity audit is not a new contest that entitles us to promote the secondary model. Its behaviour across the earlier studies remains mixed, and the present scenario family was examined because the uncertainty was already known.
The proper use of the model comparison is diagnostic: it shows which claims and procedures depend on the disputed feature. It does not confer authority to decide that feature.
9. What was executed and verified
We verified all 50 entries in the retained Volume 13 package’s checksum manifest and all 55 entries in Volume 15’s. The exact model file matches between them. The source-index and summary files copied into this audit are separately hash-bound.
A new standard-library implementation groups the event-level union and calculates the fixed conditional scores directly from the binary-feature counts. For each scenario, it also calculates each batch and page separately. The largest disagreement between the union score and the sum of the compatible batch contributions is below 7 × 10−16. The largest difference from the saved parent score is below 2 × 10−15. [3]
Thirty-two unit and regression tests passed. They check duplicate rejection, policy separation, page and group overlap, malformed features, source offsets, input tampering, frozen coefficient handling, exact small-case arithmetic, the two-event source delta and the distinction between zero and no evaluation information.
A separate numerical path uses statsmodels’ conditional-likelihood implementation at the fixed coefficients, with no fitting. Across the union, component and page checks it agrees in 114 numerical comparisons and 48 missing-score-state comparisons. The maximum numerical difference is approximately 9.8 × 10−15. These repeated views of shared data are implementation checks, not 162 independent experiments. [3, 5]
The first execution of the audit stopped because it applied the richer evaluation-record schema to an older training metadata profile. That parent profile has no separate policy or character-offset fields. We added a distinct metadata validator for its limited role in checking group exposure; we did not invent missing offsets or rescore the training rows. A regression test retains that distinction.
The design is retrospective, the input packages were already known, and both implementations were prepared within this investigation. The check establishes reproducibility at the retained event-index layer. It does not independently verify the original transcription or provide an external scientific review.
10. The unresolved image question does not disappear
The image-access work in this continuation did not yield a sufficiently detailed, retained, provenance-bound pixel object for the terminal-stroke decision. The f88r photographic overview is useful for orientation, but it does not close the high-resolution reading requirement established in Volume 16.
Different routes to a source can provide complementary information. A catalogue can identify the leaf, a manifest can identify an image service, and an interlinear display can document competing readings. None of those roles makes a metadata pointer or rendered transcription an image-level adjudication.
The twelve-target review packet described in Volume 16 therefore remains the appropriate next visual task: locate the running-text occurrence rather than the same-spelling label, retain the predecessor and gap, inspect the four prefixed comparators, and assess the proposed reference terminals without treating their transcription labels as answers. [1]
An inability to see the stroke reliably is not an observed mismatch with either candidate. Nor does an inaccessible route establish that the image is unavailable everywhere. This run’s access limitations are recorded as limitations of this run.
The combined positive score does not reduce that obligation. Resolving the reading would establish a fact about the source that the aggregate calculation cannot provide.
11. The correction record now has a defined reach
The local claim is qualified: its sign depends on an unresolved, source-attested ending. The combined diagnostic is also qualified: it remains positive under the tested alternatives, but it is a retrospective sum over selected, dependent observations.
The source verdict remains open. No hypothesis about language, cipher, notation, copying or structured generation receives a semantic promotion. The separate current-final-y and blind visual-component studies are not completed by this calculation.
The primary historical archive remains 1,031 events. The scenario files are not appended as extra observations, and the frozen training population remains 685. No earlier article is rewritten to imply that this scope analysis was known when it was published.
For later work, the next useful choice is between obtaining genuinely different source evidence and explicitly designing a broader reading-uncertainty audit. Another aggregation of the same favourable totals would not answer the remaining image question.
A sound correction should be neither smaller nor larger than its evidence. Here it reverses one local result, reduces a compatible total, and leaves the glyph unresolved. Keeping those three facts separate is the advance recorded in Volume 17.
Sources and reproducibility record
[1] Immediate predecessor. Vol. 16 — The Source-to-Image Check, WordPress post 154303. Its published source-address and review-packet account was read before continuing. This audit does not claim to have completed its image-review task.
[2] Reading scenarios. Vol. 15 — The Reading-Sensitivity Check, WordPress post 154271, and its retained scenario indices derived from Volume 14. The l/y alternatives and affected event identities are inherited findings; their existence is not a new transcription discovery here.
[3] Executed local audit. EDKSG-VOY-V017-CORRECTION-SCOPE-01. The accompanying package contains the retrospective design, byte-bound parent inputs, fixed models, standard-library scorer, scenario and stratum ledgers, tests, numerical comparison and publication record. No plaintext transcription corpus, new image verdict, model fitting or independent review is supplied.
[4] Compatible evaluation batch. Vol. 13 — The Frozen-Coefficient Check, WordPress post 154267, and its retained event indices and frozen-model file. These form the unaffected evaluation component, not new source observations acquired for Volume 17.
[5] Conditional-model reference. statsmodels: ConditionalLogit, consulted 5 September 2026. Used for the conditioning interpretation and the separate numerical likelihood check. Its intercept-removing conditional model does not establish the historical independence of our source observations.
Catalogue continuity: Vol. 3A — The Extraction Gate and Vol. 3B — The Source Identity Cross-Check retain their original URLs and historical identifiers. This work follows the already-published Volume 16. Earlier articles and the public hub are preserved.
Collection: Voynich Research Library · Editorial framework: Wintour House.
EDKSG-VOY-V017 · Edition 1.0.0 · 5 September 2026 · Baseline EDKSG-VOY-B000. Retrospective correction-scope audit. Image adjudication, complete parent-source reconciliation and independent visual assessment remain open. No semantic decipherment is claimed.