Voynich | edkSG Research Volumes
Volume: EDKSG-VOY-V004 · Edition: 1.0.0 · Edition date: 5 September 2026, Singapore
Record type: executed geometry-preservation and analysis-eligibility pilot on selected transcription excerpts.
Previous: Vol. 3 — The Extraction Gate · Voynich Research Library
Our earlier parser stopped at f68r1.5.
The displayed record was short. It had an address, a text-location type and a transcribed payload. What it did not have was the paragraph-start context required by our previous pilot. The programme refused to continue.
That refusal exposed the next problem. A record can be worth preserving even when it cannot yet enter a particular experiment. If we make experimental eligibility a condition of preservation, the research collection can begin losing precisely the material that challenges its assumptions.
This volume separates those decisions. In a new sample of 85 text-location records, the revised preserving reader retains all 85 while identifying three P-type records without recovered paragraph markers. It does not invent the missing paragraph. It keeps those records outside the marked-paragraph comparison. [2, 3]
The geometry check also changes how neighbouring records are treated. Joining each record to the next within the same page produces 82 candidate links. Only six meet our declared rule for a link inside an explicitly marked paragraph. The other 76 are set aside for that analysis—not declared historically meaningless or unrelated. [3]
We have improved the distinction between what the transcription contains and what a particular test is entitled to use. No q/y effect, language identification, visual-component result or decipherment is reported here. Complete upstream-file byte verification remains open.
Reading route: the new sample · the failed assumption · 82 candidate links · the circular-order check · next research step.
1. The next step begins where the previous test was narrowest
The Extraction Gate tested 106 selected records and demonstrated a specific destructive-cleaning failure. Its sample contained paragraph-text and label records, but no circular- or radial-text records. The article explicitly left broader geometry coverage and full-corpus validation unfinished. This volume addresses that declared coverage gap. [1]
We did not replace the earlier source excerpts, rewrite their results or treat their 32 software tests as evidence that every manuscript layout had already been handled. We retained the old implementation as an identifiable reference and tried it on new, deliberately different material.
The new material comes from the displayed Zandbergen–Landini transcription. The work is still one layer away from a verified raw-file acquisition and several layers away from the manuscript itself. Our local excerpt preserves a transfer from a browser display. It does not carry an authenticated claim to the original server’s exact bytes.
Fresh direct-download attempts failed again. A repository catalogue supplied metadata for a possible copy, but no complete byte-preserving transfer was established through that route. We record that as an access limitation, not a checksum mismatch, a changed source or a successfully acquired corpus.
That leaves room for bounded work without relaxing the acquisition requirement. We can investigate whether our own processing preserves selected records and distinguishes their uses. We cannot call the outcome a reconstruction of the complete historical computational input.
The relevant question for this checkpoint is therefore narrower: does the preserving procedure still work when the page contains several kinds of text arrangement, including material that does not satisfy our preferred paragraph model?
2. Three page units, four kinds of text location
We selected f67r1, f67v1 and f68r1, transferring their displayed page headers and complete numbered locus ranges into a separately named local excerpt. The selection was purposive: it was made to exercise circular and radial location classes absent from the previous pilot. It is neither a random manuscript sample nor a blind scientific evaluation set. [2, 3]
A locus is an addressable text location in the transcription. The term does not, by itself, establish a sentence, a linguistic word or an independently observed physical line. The table below counts records in our local transfer.
| Page unit | P | L | C | R | Total |
|---|---|---|---|---|---|
| f67r1 | 4 | 0 | 3 | 12 | 19 |
| f67v1 | 0 | 12 | 0 | 17 | 29 |
| f68r1 | 7 | 29 | 1 | 0 | 37 |
| Total | 11 | 41 | 4 | 29 | 85 |
The three per-page family counts and the recorded paragraph-start totals agree with the inherited matrix. That is a consistency check between related representations, not independent validation of the transcription. The matrix derives from the same upstream source family. [3, 5]
The sample’s dependence is unusually important. In the inherited matrix, all three page units belong to the same recorded bifolio group, Q9-B1. All carry effective hand 4, the M2 visual classification and an unassigned Currier field. These are inherited assignments, not classifications newly established from images in this run. [5]
We therefore have 85 records across varied encoded geometries, but not 85 independent manuscript examples, three independently sampled production environments or a representative estimate of the whole book.
This connects directly to Volume 2. The A/B-only selection examined there excluded the M2 population. Here we begin processing some of that previously excluded material without pretending its Currier classification has become known. We expand representation coverage while retaining the reason it may need a different analytical treatment.
3. Preservation and eligibility were doing the same job
The earlier implementation required every P-type record to sit inside recovered paragraph-start and paragraph-end context. That was an explicit restriction of a bounded pilot. It worked for its selected development material; it was not a certificate of complete format coverage. [1]
Applied unchanged to the new excerpt, it stops with the message that a P locus lacks recovered paragraph-start context. Inspection locates the first such record at f68r1.5. The displayed block contains three records, f68r1.5–7, typed Pb without the paragraph markers required by that pilot. [2, 3]
The useful response is not to add markers until the file passes. That would alter the representation to satisfy the programme. Nor should we silently delete the three records and report the remainder as a complete reading of the selected source.
Instead, the new preserving reader retains the payload, address, type and other supplied information. It records that marked-paragraph context has not been recovered. A separate eligibility function then refuses to use those records as if their paragraph membership had been established.
This produces a small but consequential architectural change: “preserved” no longer implies “admitted to this experiment”.
The distinction does not lower standards. It prevents a narrow experimental standard from being misapplied as a reason to erase evidence. Unsupported syntax still causes an explicit stop; a recognised record lacking the context needed for one particular analysis is retained with that limitation.
We do not decide in this volume whether the absent markers reflect an omission, a transcription convention or another circumstance. The evidence recovered here supports a statement about the supplied record and our processing rule. It does not settle the reason behind the record’s form.
4. Eight marked records, three unmarked records, seventy-four others
The revised reader assigns three descriptive states in this local audit. Eight P records belong to explicitly marked paragraph spans. Three P records lack that recovered context. The remaining 74 records are not P-type locations. These states are analysis bookkeeping, not proposed classes of medieval meaning. [3]
| Recovered context | Records | Treatment in this pilot |
|---|---|---|
| Explicitly marked P paragraph | 8 | Retain; consider for the declared paragraph-link rule |
| P type without recovered paragraph markers | 3 | Retain; do not manufacture paragraph membership |
| Non-P location | 74 | Retain; keep its geometry distinct |
The preservation record retains a null paragraph identity for the unmarked block. Null does not mean “no paragraph ever existed”. It means that this procedure has not recovered one from the supplied markers.
This wording matters because a database field can quietly become an argument. If a field says “not a paragraph” when the evidence merely says “paragraph membership not recovered”, later software may exclude a live possibility without the researcher noticing.
The same caution applies to the 74 non-P records. They are outside the present paragraph comparison because of the question we are asking and the source types we retained. Their exclusion does not establish that they lack grammar, sequence, information or a relationship with the surrounding material.
A future label or diagram-text investigation can ask a different question of those records. It should reuse the preserved representation, not reconstruct it from a prose-only dataset that has already discarded them.
The immediate result is modest: we now know which exclusion belongs to the analysis rather than to the source.
5. What happens to the 82 neighbouring-record links?
For each selected page, we paired each record with the next record in file order. We did not cross page boundaries. The three pages therefore supply 18, 28 and 36 candidate pairs: 82 in total. This is the deliberately broad candidate generator against which the narrower rule is compared. [3]
The admission rule is our stated pilot design. Both records must be P-type, share the same recovered marked-paragraph identity, have consecutive locus numbers and satisfy the permitted relative-location conditions. A paragraph ending or a new paragraph beginning prevents admission as an internal paragraph link.
The result is six admitted pairs, three in the marked paragraph on f67r1 and three in the marked paragraph on f68r1. Seventy-three candidates involve a non-P location. Three fail because marked-paragraph context is missing. These are mutually exclusive first-reason dispositions in the saved audit, not estimates of three natural manuscript categories. [3]
| Disposition under the declared rule | Candidate pairs |
|---|---|
| Admitted within a recovered marked paragraph | 6 |
| Set aside: at least one non-P location | 73 |
| Set aside: missing marked-paragraph context | 3 |
| Total within-page candidates | 82 |
This is not a discovery that 76 links are historically false. It is a calculation showing that file adjacency supplies a much broader candidate population than this one declared paragraph experiment admits.
Conversely, admission does not prove that the six remaining pairs are linguistically connected in the way a later model assumes. It establishes only that they meet the recorded source-context rule. Token uncertainty, the target contrast and the interpretation of a line transition still need their own checks.
Keeping those two limits visible prevents either side of the result from becoming too strong. Exclusion is not proof of irrelevance; inclusion is not proof of meaning.
6. The plus sign does not turn two circles into prose
The format’s locator and its text-location type answer different questions. The locator describes a relative position; the type describes the kind of text arrangement. In the specification, plus indicates a generally lower position, while the Cc type identifies clockwise circular text. These codes do not establish a sentence continuation. [4, Tables 8–9]
The new excerpt makes that distinction concrete. At f67r1, circular records 6 and 7 carry a plus locator. The audit preserves it, but does not admit either neighbouring circular-record pair to the marked-paragraph comparison. [2, 3]
A programme that uses plus alone as a universal instruction to concatenate would have lost the type distinction. A programme that ignores plus altogether would lose positional information that could be useful for another question. The safer first step is to retain both fields and make their analytical use explicit.
Likewise, the selected f67v1 label sequence includes ampersand locators. The specification distinguishes their circular spatial relationship from ordinary same-line positioning. That is useful metadata; it is not enough to assign lexical adjacency or a universal reading path. [4, Table 8]
The principle is easy to express but demanding to implement: a field is not allowed to acquire more meaning merely because it is convenient for the next calculation.
For our pathway, the concrete repair is to make the candidate link depend on the declared combination of fields. The source preserves geometry. The experiment says which geometry it uses. Neither is silently substituted for the other.
7. The radial records remain separate
The sample contains twelve Ri records and seventeen Ro records. The specification uses those subtypes to distinguish inward and outward radial text. The audit retains the supplied subtype and the record boundary. It does not reverse the character string in response to that label or join one radius to the next. [2–4]
That last restraint matters. An outward orientation in a drawing is not an instruction to reverse a transliteration already supplied in reading order. Reversing it would create a new analytical transformation, which would require its own justification and separate record.
Nor does a sequence of radii in a file automatically identify the historical order in which a reader moved between them. The sequence may be a way of listing the locations. The next radial item can be nearby in the transcription while belonging to a different visual path.
We therefore retain radial records for a future geometry-specific analysis, but make no cross-radius transitions in the present paragraph rule. This is a deliberate restriction of the experiment, not a claim that cross-radius relationships cannot exist.
A later proposal could define a path, state the image evidence for that path and compare its predictions against alternatives. It would then have something testable. Starting with one flattened stream would hide the path choice before the comparison began.
No manuscript-image inspection of those paths was conducted for this volume. The present evidence is the displayed transcription and the encoded distinctions preserved from it.
8. A circle needs a different sensitivity check
The four C-type records raise another issue: what happens when a circular arrangement is represented as a linear sequence?
We tested a deliberately explicit mathematical model. Using the previous pilot’s diagnostic segmentation with doubtful seams split, each selected circular record becomes a finite list of spans. Alternatives and ligature constructs remain opaque; these spans are not asserted to be linguistic words. [3]
Under a closed-cycle model, the last span is connected back to the first. We compared the multiset of directed adjacent pairs under every cyclic rotation of each list. Rotating changes which span is written first in the list without reversing its orientation or changing which spans neighbour one another in the assumed cycle.
The four lists contain 37, 41, 31 and five diagnostic spans. All 114 constructed rotations preserved their respective closed-cycle pair counts. [3]
| Circular record | Spans / rotations checked | Closed-cycle pair occurrences | Open-chain pair occurrences |
|---|---|---|---|
| f67r1.5 | 37 | 37 | 36 |
| f67r1.6 | 41 | 41 | 40 |
| f67r1.7 | 31 | 31 | 30 |
| f68r1.37 | 5 | 5 | 4 |
This result is a property of the model and its implementation. For a nonempty list of n spans, a closed cycle has n directed neighbour occurrences. An open chain has n minus one. Rotating a closed cycle leaves the same pairs with the same multiplicities. The arithmetic does not establish that a historical ring should be treated as closed.
The distinction is critical. These are not 114 independent pieces of evidence that Voynich is cyclic, nor an experiment discovering the author’s reading rules. They are constructed transformations of four source-derived lists, used to check the consequences of an explicit assumption.
The model’s usefulness is diagnostic: a future analysis claiming not to depend on where a closed ring is cut should survive this rotation test. If it does not, the result depends on something other than the invariant closed-cycle neighbour structure.
9. A recorded starting point may still matter
Three of the selected circular payloads carry the inline comment 00:00; the fourth carries 08:30. These source annotations are preserved. The rotation diagnostic changes a derived list, not those source strings. [2, 3]
The exercise does not authorise us to discard an original opening, a gap, a marker or a conventional starting position. A physical ring can have a meaningful entrance. If there is evidence for one, the starting point becomes a feature to investigate rather than an arbitrary inconvenience to remove.
Two different questions follow. One asks which result remains unchanged when the analytical cut moves. Another asks whether the recorded opening has a special relationship with the text. Those questions should not be combined into a test that first assumes an arbitrary opening and then announces that the first span is special.
For example, an open-chain representation omits one closing pair occurrence. Which occurrence is omitted depends on where the cut is placed. That is not necessarily a mistake: an open-chain model may be appropriate. It is a choice whose effects need to be visible.
The present implementation can compare these model consequences without choosing a historical interpretation. Recovering that interpretation would require additional evidence, potentially including the image, the placement of gaps and the relationship between the written ring and surrounding elements.
The new ability is not to make all circles equivalent. It is to stop an unnoticed linearisation choice from masquerading as an observation.
10. What was checked after the change
The revised preserving reader reproduced all 85 local payloads from its event strings. The saved event offsets point back to the recorded substrings. Reassembling the retained local lines reproduces the local excerpt bytes. These are checks against the transferred working file, not the original server file or the manuscript image. [3]
We also ran the new preserving reader against the earlier 106-record excerpt. The comparison found no differences in the eleven selected preservation fields: locus identity, page, number, locator, type, payload, header, effective metadata, events and the two paragraph-marker flags. The new reader adds explicit context information without changing those checked inherited fields. [3]
The earlier implementation’s 32 tests were rerun successfully. Twenty tests for the new implementation also passed. They exercise unmarked-record preservation, exclusion from the marked-paragraph view, retained geometry, paragraph and page boundaries, malformed-input rejection, local reconstruction and the circular-model properties. [3]
The two test groups have different meanings. Rerunning the old suite checks that the archived implementation still behaves as recorded. Testing the new implementation checks its newly declared behaviours. Neither count is a probability of correctness or a measure of how much of the manuscript has been solved.
A narrow second counting path agreed on 85 records, the four location-family totals and two paragraph starts paired with two paragraph ends. Both implementations and the second counting path were prepared within this investigation. We do not present them as independent expert review.
The new sample was inspected during development and prompted the change in handling. It is therefore development evidence, not a held-out scientific replication. The next evaluation must preserve that distinction.
11. What the change still refuses to do
The preserving reader is still a bounded merged-EVA implementation. It is not a general certificate of compliance with every IVTFF feature. Unsupported constructions remain explicit limitations, including the broader treatment of wrapped and interlinear records.
Malformed brackets, duplicated identities, inconsistent metadata changes and unpaired explicit paragraph markers do not become acceptable simply because we have separated preservation from analysis. The new rule is not “accept anything”. It is “retain recognised information while identifying the context that is still missing”.
There is also no automatic promotion from a recognised record to a meaningful word. The event representation retains modern transcription labels. It neither assigns sounds to them nor resolves competing readings according to which one makes a preferred interpretation work.
A programme could preserve every byte and still organise its analysis incorrectly. That is why round-trip checks, context checks and eligibility checks are separate. Each catches a different kind of failure; none substitutes for the others.
For a later q/y analysis, the six admitted cross-locus links are not a sufficient experimental population. They are examples of a record-level rule. The target token definitions, treatment of uncertainty, within-locus transitions, source completeness and evaluation design remain to be fixed for the actual statistical question.
The work reported here makes those future choices more inspectable. It does not quietly execute them under the heading of a parser upgrade.
12. The research-state change
At the end of The Extraction Gate, the preserving procedure had been exercised on paragraph-text and label material, with a narrow paragraph-context restriction. It had not been tested on circular or radial records.
At this checkpoint, an additional 85-record geometry sample has been examined. A real scope limitation in the older pilot has been exposed. The new reader preserves the affected block while keeping its experimental ineligibility explicit. A complete ledger records all 82 within-page candidate links and their dispositions. A separate circular-model check makes the effect of the assumed cut and closure visible. [3]
The gain is not another theory of the manuscript’s purpose. It is a more accurate description of the population on which a future theory would be tested.
This is also why the result belongs in a new dated volume rather than a rewrite of Volume 3. The earlier pilot’s limitations were part of its record. We have now encountered one of them, changed a particular handling rule and checked the consequences. The old work remains a recoverable predecessor.
The local research identifier for this checkpoint is EDKSG-VOY-V004-GEOMETRY-01. It identifies this run, not a new production-system identity or a source-owner certification. The earlier q/y and visual-test identities are unchanged.
13. What remains unresolved
Complete upstream-byte acquisition remains open. The existence of a copy, a filename, a matching version label or a published checksum record does not mean this run has received and hashed the complete bytes. We have not computed the primary source’s SHA-256 in this checkpoint.
Full-format acceptance is also open. Passing an 85-record extension does not establish that all 5,385 locations represented in the inherited matrix can be read and classified correctly by the current implementation. Every additional supported construction needs its own accounting and tests.
The source-to-image comparison has not been performed for these new records. Consequently, the conclusions concern the transcription and the research procedure. They do not certify the physical page’s actual reading paths, whether any ring is continuous, or whether the assigned text-location type captures every relevant feature.
No original boundary-effect event table was reconstructed. No new q/y association was estimated. The blind recurring-component experiment was not run, and the text and visual lanes were not combined. No language, function, origin or translation was identified.
Those limits do not erase the result. They define it. We have learned how to preserve more of the represented manuscript without forcing every part through the same experimental entrance.
14. The next justified research step
The next corpus-level task is still to obtain a byte-preserving copy of the intended transcription, or to declare and retain a different input edition explicitly. That source must be reconciled against the inherited manifest before an exact old-input replication is claimed.
In parallel, the representation test now has a clearer contract. Recover every included location, preserve its payload and metadata, account for unsupported constructions, and separate preservation status from each analysis-specific admission rule. Do not fill missing context by inference without recording that inference as a different layer.
The next statistical protocol should name one population and one contrast. It should not pool marked paragraphs, unmarked blocks, circular records, radial records and labels merely because a parser can now read them all.
Where a control excludes a population, report the exclusion. Where a geometry requires a different path model, state the model. Where the data leave two interpretations open, retain both rather than choosing whichever supplies the desired statistic.
A useful future result might strengthen a boundary hypothesis, weaken it or reveal that the current sample cannot distinguish its explanations. We do not need to prescribe the outcome now. We need to make the route to that outcome reproducible.
The advance in Volume 4 is a distinction: preserve the record first; decide what a test may infer from it second.
The manuscript has not become simpler. Our account of what we are doing to its representation has become more precise.
Sources, methods and catalogue note
[1] Exact predecessor. Voynich | edkSG Vol. 3 — The Extraction Gate, edition 1.0.0, WordPress post 154062. Its retained research package supplies the previous implementation and 106-record development excerpt.
[2] Selected primary transcription display. Zandbergen–Landini ZL3b-n, displayed version 3b, 13 May 2025. This run transferred the page headers and loci f67r1.1–19, f67v1.1–29 and f68r1.1–37 from the browser display on 5 September 2026. Standalone modern commentary was omitted; inline annotations were retained. This is not a raw-file byte snapshot.
[3] Executed local research record. EDKSG-VOY-V004-GEOMETRY-01. Retained materials include the selected excerpt, preserving reader, all 82 candidate-link decisions, circle-model outputs, test logs, comparison against the inherited sample and run manifest. The reported numerical results come from that run. Its source transfer is manual, its sample is purposive, and independent scientific replication is not claimed.
[4] Format authority. René Zandbergen, IVTFF specification, document issue 2.0.2, 8 July 2025, particularly Tables 8–9 and section 6.9. Relevant page images were inspected. The specification defines representation conventions; the present admission rule and cycle diagnostic are our declared analytical choices.
[5] Inherited coverage and grouping. Vol. 2 — The Input Audit and its retained matrix rows, derived from the frozen Voynich_Test_1b_Frozen_Segmentation_Matrix.xlsx. The page groupings and classifications remain attributed to that inherited representation.
Catalogue disambiguation. At this check, two published works carry the Volume 3 label: The Extraction Gate (post 154062) and The Source Identity Cross-Check (post 154068). This volume continues the former by exact title and URL. It records the shared-label ambiguity without renumbering or rewriting either earlier work.
Collection: Voynich Research Library · Series starting point: Vol. 1 — The Research Baseline · Editorial framework: Wintour House.
EDKSG-VOY-V004 · Edition 1.0.0 · 5 September 2026 · Baseline EDKSG-VOY-B000. Earlier editions remain unchanged. Corrections should identify the affected claim, source representation and edition. No independent scientific replication or decipherment is claimed.