Voynich | edkSG Research Volumes · EDKSG-VOY-V013 · Edition 1.0.0 · 5 September 2026, Singapore
Research status: executed frozen-coefficient comparison on three additional source-identified page samples. The coefficients were fixed before these source files were retrieved in this run. This is not a certified historically unseen test, independent scientific review, complete-corpus analysis or decipherment.
The coefficients were already written down when the next pages arrived.
That is the difference this volume introduces. We did not observe the new result and then adjust the model until it described that result. We fitted the common and partial-sharing procedures to the earlier 685-event collection, recorded their parameters, and only then retrieved the three selected source files.
The frozen common model improves the conditional score against zero on all three additional page samples. Its total improvement is +5.693 across 73 outcome-varying evaluation observations. The pages contribute very unequal information: f34r supplies just two of those observations, while f77r supplies 34 and f105r supplies 37. [2]
The partial-sharing model scores slightly better under the primary split policy, at +5.873. But that small ordering reverses when doubtful seams are joined or when pairs touching them are excluded. There is no stable victory for the added hand-specific variation across these declared views. [2]
The structural clue now survives a bounded test with its coefficients fixed before the new source acquisition. The hand-specific refinement still does not earn a robust preference. Neither result tells us what the writing means.
1. What was fixed before the new source was read
Volume 12 ended with a useful restraint: do not keep searching for a penalty that makes already-exposed pages look convincing. The next step should add evidence or a justified measurement, rather than another way to fit the same observations. This volume follows that instruction. [1]
The primary estimator is the common ridge model from that comparison. It has one coefficient for whether the preceding represented span ends in y. The secondary estimator is the previously declared partial-sharing model with its hand-deviation penalty fixed at one. The common-component penalty is 0.25 in both.
Both were fitted once to the existing 685-event split-policy training collection. The fitted common coefficient is approximately 1.544, corresponding to an odds-ratio equivalent of 4.683. The secondary model supplies equivalents of 8.007 for recorded hand 2 and 4.179 for recorded hand 3, the assignments encountered in the new samples. These are fixed model parameters, not odds ratios re-estimated from the new pages. [2]
The selected pages, source-column choice, primary and secondary procedures, three spacing views and no-refitting rule were recorded before the target files were fetched. The local design and frozen-model records have content hashes. They identify the decision being tested; the hashes are not proof of scientific truth or independently witnessed preregistration.
Only two fitted procedures enter this comparison. We do not repeat the earlier five-candidate search, choose a different model for each new page, or use the new results to tune the penalty. The zero-coefficient reference remains the same conditional comparison with the predecessor association removed.
This is stronger separation within the present run than another retrospective fit to all available rows. It is still not a claim that nobody in the wider research programme has ever examined these pages. The model family and research question were developed from earlier work, and the selection is purposive.
2. Three new samples, not three new manuscripts
The inherited metadata frame supplies the sampling descriptions. We selected f34r, f77r and f105r to revisit the three broad contexts examined in Volume 9 while moving to different recorded bifolio groups. Their text outcomes were not used to select them in this run.
| Page unit | Recorded group | Recorded hand | Inherited class | P / L records |
|---|---|---|---|---|
| f34r | Q5-B2 | 2 | M1 | 15 / 0 |
| f77r | Q13-B3 | 2 | M3 | 40 / 10 |
| f105r | Q20-B3 | 3 | M6 | 37 / 0 |
The three recorded groups are disjoint from the nine groups in the fitted event collection. No training event or directly sampled training page reappears in the target collection. The check concerns the supplied grouping; it does not independently reconstruct the binding or establish statistical independence between groups.
The shared features matter as much as the differences. The samples belong to one manuscript and one transcription lineage. They share a Currier classification, and two share a hand assignment. They are not a balanced experiment separating visual class, scribal practice and textual population.
The M1, M3 and M6 labels retain their earlier descriptive roles. They do not authorise us to rename the samples as three decoded subjects. A difference between their scores would be an observation to explain, not proof that the pictures caused the difference.
All three samples remain in the account, including the sparse first one. A page that yields few useful comparisons is not replaced after inspection by a page with a more attractive result.
3. The source objects are exact; their historical readings are not certified
Direct retrieval of the complete origin-host file and the raw GitHub mirror again failed at name resolution in the working environment. The failure is recorded as an access outcome, not as a source mismatch. No full-parent checksum verification is claimed.
The connected repository route returned three complete diplomatic text objects instead. Each includes location identifiers, location types, a raw-IVTFF column and a cleaned column. Locally reconstructed files were admitted only after their computed Git blob identities matched the identities returned by the connector. All three matched. Separate SHA-256 digests identify the local files. [2, 3]
The cleaned columns were reconstructed only to verify complete-object identity. They do not enter the analysis. The raw columns supply the preserving parser, so a cleaned string cannot silently replace an uncertain reading or join text around a drawing interruption.
In total, the source objects contain 102 location records: 92 P-type records and ten labels. The labels belong to f77r. They are preserved, but excluded from this paragraph-text experiment. The page totals and the two, three and ten paragraph-start counts agree with the inherited frame. [2]
Every retained raw payload can be reconstructed from its parsed events, and its recorded offsets return the expected substrings. These are checks of the acquired representation. They do not establish that every mark on the manuscript was transcribed correctly.
The distinction is deliberate: content identity secures the computational handoff, while source-to-image accuracy remains a separate evidential question. Three exact derivative files are not the complete origin file and not three independent transcriptions.
4. The representation does not change to suit the new pages
The extraction, geometry-preservation and boundary-analysis modules are reused without changes. Their content hashes are checked against the inherited copies. A new adapter transfers the source raw fields into the same representation expected by those modules.
Under the primary policy, doubtful seams divide analytical spans, but an admitted transition must cross a confident dot separator. Drawing interruptions remain barriers. Alternative readings, ligatures, uncertain marks, extended codes and attached annotations remain reasons to exclude an affected plain-span pair. Exclusion does not delete the span and connect its neighbours.
The target remains oR versus qoR, where R is a nonempty remainder. Removing the optional leading q defines a matching string, not a historical morpheme. The preceding y is a represented final character, not an assigned suffix or instruction.
The 92 eligible location records supply 757 candidate pairs. Of these, 269 enter the oR/qoR comparison, 380 are outside that contrast, 81 are blocked by doubtful or drawing-related boundaries, and 27 involve nonplain or annotated spans. These are mutually exclusive first-reason dispositions, and they sum to 757. [2]
The resulting 269 observations are not 269 independent experiments. Adjacent pairs can share spans, repeated forms can recur within one page, and the pages have related provenance. The candidate and admitted ledgers retain their addresses instead of allowing an aggregate count to conceal those relationships.
5. What a frozen conditional score actually asks
The evaluation groups events by exact page, recorded hand and current-form family. It asks how well the fixed coefficient accounts for the allocation of qo outcomes given each evaluation stratum’s observed total number of qo outcomes.
That condition is essential. Conditional logistic modelling removes stratum intercepts through conditioning. It is not an ordinary system that predicts every unread token without knowing the target outcome totals. We retain the narrower interpretation already made explicit in Volumes 11 and 12. [4]
For both frozen models, we subtract the conditional log-likelihood at coefficient zero from the conditional log-likelihood at the fitted coefficient. Positive means the fitted association assigns more conditional probability to the observed allocation than the zero reference does. Penalties affect training only; they are not subtracted from the evaluation score.
The new split-policy sample has 73 outcome-varying observations. Sixty-two, distributed across fifteen strata, also have predecessor variation and directly inform the coefficient comparison. The remaining eleven contribute constants that cancel in the model-versus-zero difference. Both models are scored on exactly the same 73 observations. [2]
No coefficient is fitted to these target outcomes. They determine the observed allocation and the conditioning totals, not a new estimate. This distinction gives the test a useful fixed-model meaning without pretending it is unconditional word prediction.
6. The frozen common model improves all three conditional comparisons
| Page unit | Admitted transitions | Conditional observations | Common model gain | Partial-sharing gain |
|---|---|---|---|---|
| f34r | 29 | 2 | +0.500 | +0.575 |
| f77r | 107 | 34 | +2.545 | +2.686 |
| f105r | 133 | 37 | +2.649 | +2.611 |
| Total | 269 | 73 | +5.693 | +5.873 |
The common model’s positive result is no longer confined to reassigning already-analysed pages to retrospective folds. It now extends to three additional page objects acquired after the current coefficients were fixed.
The two larger conditional samples carry most of the gain. Their contributions are similar in magnitude, so the new total is not supplied almost entirely by one of them. That is a description of this batch, not a claim that the overall programme’s unevenness has disappeared.
On the primary view, the partial-sharing model gains about 0.180 more than the common model in total. It does better on the two hand-2 samples and slightly worse on the hand-3 sample. This ordering is recorded even though the same model was less successful in the preceding retrospective comparison.
We do not force new evidence to agree with the previous volume. Equally, one small favourable difference is not enough to announce that the earlier conclusion has been overturned. The declared sensitivity checks test whether that preference survives nearby representation choices.
7. The smallest positive result is a two-observation comparison
f34r contributes 29 admitted transitions, but only one informative same-form stratum, containing two observations. Its positive conditional score must not be described as though all 29 observations independently confirmed transfer.
For this two-observation alignment, the gain at coefficient β is log(2) − log(1 + exp(−β)). At the frozen common coefficient, it is approximately 0.500. Even an indefinitely large positive coefficient could not raise this particular gain beyond log(2), approximately 0.693.
That mathematical ceiling helps interpret the result. The page supplies one small correctly directed comparison, not a broad estimate of how often the association holds throughout the page or its illustration class.
Its pooled unadjusted odds ratio is 5.00, but that table has only one non-y-to-qo occurrence. The ratio is therefore highly sensitive to small changes in the admitted count. It is retained as a description, not promoted into a stable population estimate. [2]
The same discipline applies to the three-positive-pages headline. It is not a binomial experiment with three independent equally informative trials. The page selection, data dependence and unequal support prevent that interpretation.
8. The larger descriptive table remains separate from the transfer score
The complete primary batch contains 119 y-to-qo events, 56 y-to-o events, 23 non-y-to-qo events and 71 non-y-to-o events. Its pooled descriptive odds ratio is approximately 6.560. The page-specific ratios are 5.00 for f34r, 13.33 for f77r and 4.32 for f105r. [2]
Those are estimates calculated from the new observations. The frozen common coefficient, whose odds-ratio equivalent is 4.683, was calculated from the earlier training set. Confusing the two would undo the central separation of this volume.
The pooled table also ignores the exact-form stratification used by the transfer score. It can be useful for accounting and orientation while answering a different question. A large pooled ratio does not by itself demonstrate a same-form relationship or good transfer.
We do not attach a population p-value, an independence-based confidence interval or a probability of decipherment. The sample is purposive, the events are dependent, and the source remains a modern representation of an undecoded manuscript.
The conditional improvement is more directly relevant to the frozen-model question, but its own limits remain. It evaluates one model feature under one family of strata, not every possible explanation of the writing.
9. The hand-specific advantage changes sign
The other two policies were declared before the target files were acquired. One joins doubtful seams; the other excludes a candidate pair touching one. Both retain drawing interruptions as barriers. The training coefficients stay frozen from the primary split-policy training set in every case.
| Target representation | Admitted events | Conditional observations | Common gain | Partial-sharing gain |
|---|---|---|---|---|
| Split doubtful seams | 269 | 73 | +5.693 | +5.873 |
| Join doubtful seams | 271 | 70 | +4.837 | +4.795 |
| Exclude touching pairs | 241 | 65 | +5.734 | +5.573 |
The common model remains positive on all three pages under every tested policy. The partial-sharing model also remains positive, but its advantage over the common model is not stable: about +0.180 in the split view, −0.042 in the joined view, and −0.160 in the exclude-touching view.
This is a representation-sensitivity finding, not a significance test of the difference. The denominators change with the view. We cannot attribute every score movement to a boundary choice while pretending the evaluated observations have remained identical.
The appropriate conclusion is narrower than declaring a winner: this batch supports the common structural clue under the tested alternatives, while failing to provide a consistent preference for the additional hand deviations.
The current secondary result does not erase Volume 12’s earlier negative comparison, and the earlier comparison does not entitle us to suppress today’s slight primary-view advantage. Both remain part of the record.
10. What was actually verified
The three local source files match their returned Git blob identities. All 102 raw payloads round-trip through the preserving representation, and admitted source-component offsets are checked under all three policies. Page and paragraph counts reconcile with the inherited metadata.
A second raw-text path, separate from the primary scanner and event extractor, reproduces all 269 primary events, participating strings and offsets. It shares the source admission and uses an earlier documented raw-splitting implementation; it is not an independent researcher or a second transcription.
A separate likelihood calculation using statsmodels checks both frozen models on each page and on the combined batch under all three policies: 24 numerical comparisons. The largest absolute score difference is approximately 1.1 × 10−14. No target coefficient is fitted in that check. [2, 4]
Twenty-eight new unit and regression tests passed; the 28 inherited boundary tests also passed. The new checks include source tampering, frozen-parameter tampering, training-group separation, shared evaluation populations, sparse-stratum arithmetic, duplicate rejection and execution with model fitting disabled.
These are checks of the implemented procedure. They do not certify the transcription against images, demonstrate historical independence or constitute external scientific review. The package distinguishes a source-level rerun, requiring acquisition of the listed files, from a score-level rerun using the included non-plaintext event indices.
11. The archive grows without turning evaluation into future training by accident
Appending the primary new index to the compatible historical collection gives 954 distinct admitted events. The comparison opportunity grows from 33 informative strata with 180 observations to 48 strata with 242 observations. These cumulative counts describe the archive after the test; they are not used to refit today’s frozen coefficients. [2]
Nineteen page units are now directly represented in the retained development history, including the earlier page that contributes no event to this particular contrast. Eighteen contribute events. Their recorded group closure contains 56 units of the 227-unit metadata frame; the remaining 171 have exposure not established by this audit.
Those 171 are not certified unseen. The wider programme includes work predating the numbered series. The new pages themselves are now explicitly exposed and cannot become untouched evaluation material again by being copied into a differently named file.
We also keep the score histories separate. Adding this batch’s +5.693 to a previous cross-validation total would combine different fitting and evaluation arrangements. The number would not describe one coherent new experiment. Accumulated events may be counted; incompatible evaluation designs should not be collapsed into an impressive composite score.
12. What this lets us understand
The simplest defensible positive statement is that a common predecessor-y coefficient learned from the earlier collection improves the conditional allocation score in these three additional represented contexts without refitting. The improvement survives the three declared target-space treatments.
That strengthens the case for treating the relationship as a structural lead worth investigating. It does not establish one uniform law, because earlier counterperforming contexts remain and the present samples are small and selected.
The proposed hand-specific refinement receives a much less decisive result. Its small primary-view gain is sensitive to the representation, so the added complexity has not earned a general preference. This is not proof that scribes are irrelevant or that every hand-aware model must fail.
The historical explanation is still open. Linguistic dependence, encoding, copying, a state-dependent production process, transcription choices or interacting mechanisms may account for different parts of the pattern. The present model does not discriminate all those possibilities.
Nothing in the result supplies a translation of q, y or the matched forms. The current-token final-y experiment remains a different question, and the blind visual-component investigation is not completed by a textual score.
13. The next evidence question
The next useful expansion should challenge where the fixed relation fails as deliberately as where it succeeds. A new sampling rule needs to be stated before inspecting its outcomes and should broaden the contextual comparison, not simply seek more pages resembling the current positive examples.
Source-to-image checks would add a different kind of evidence. They could examine whether influential admitted or excluded transitions depend on ambiguous transcription or gap decisions. Such an audit should include both favourable and unfavourable cases rather than selecting only the pairs that support the model.
Complete parent-source acquisition remains another unfinished task. It would enable a wider coverage census under one representation. It would not automatically supply independent validation, and repeated inability to download it must not be mistaken for completion.
The present achievement is a more credible boundary between learning and evaluation inside a recorded run. The selected coefficients were fixed; the additional sources were then acquired; their actual outcomes were retained without repair-by-refitting.
The clue travelled to three more samples. Its meaning did not travel with it. Our next job is to explain the structural relationship while keeping that distinction intact.
Sources and reproducibility record
[1] Exact predecessor. Vol. 12 — The Partial-Pooling Check, WordPress post 154256, edition 1.0.0. Its 685-event training index and two declared primary model forms are reused. The preserving extraction lineage runs through Volumes 5, 7 and 9; no earlier research body is changed here.
[2] Executed local record. EDKSG-VOY-V013-FROZEN-COEFFICIENT-01. The research package contains the pre-acquisition design, fixed parameters, input and code identities, source manifest, hashed event indices, score and coverage ledgers, tests and numerical cross-check. Full third-party diplomatic files are not redistributed; their acquisition instructions and exact identities are supplied.
[3] Acquired source objects and frame. Zandbergen–Landini-derived diplomatic records in cipher_benchmark: f34r, f77r and f105r, retrieved 5 September 2026. The 227-unit frame is retained from Volume 8, whose metadata derive from the earlier frozen matrix.
[4] Conditional-model reference. statsmodels: ConditionalLogit, consulted 5 September 2026. Used for the conditioning distinction and as a separate numerical likelihood implementation, not as evidence about manuscript meaning.
[5] Evaluation boundary. scikit-learn: Cross-validation, evaluating estimator performance, consulted 5 September 2026. Its general separation of fitting from evaluation informs the design; it does not certify this manuscript’s historical group independence.
Catalogue continuity: Vol. 3A remains The Extraction Gate and Vol. 3B remains The Source Identity Cross-Check. Their original URLs and historical identifiers are preserved.
Collection: Voynich Research Library · Editorial framework: Wintour House.
EDKSG-VOY-V013 · Edition 1.0.0 · 5 September 2026 · Baseline EDKSG-VOY-B000. Fixed-coefficient exploratory extension on an identified representation. No certified historically unseen test, independent review, complete-source verification or semantic decipherment is claimed.