Voynich | edkSG Vol. 3A — The Extraction Gate

Voynich | edkSG Research Volumes
Volume: EDKSG-VOY-V003 · Edition: 1.0.0 · Edition date: 5 September 2026, Singapore
Record type: executed extraction pilot on selected browser-transferred excerpts. Full source-file verification remains open.

Previous: Vol. 2 — The Input Audit · Voynich Research Library

We deliberately gave the text to a careless cleaner.

Its instruction was simple: remove everything inside angle brackets. The result looked tidier. But across our selected excerpts, the cleaner erased 37 recorded drawing boundaries in 28 text locations. In places where the representation had explicitly separated two stretches of writing, the cleaned version joined them. [5]

Nothing about the manuscript had changed. Our software had changed the evidence available to the next calculation.

That is the central result of this volume. Before asking whether a model understands Voynich, we need to establish whether the model is receiving the distinctions that the transcription actually records. A plausible-looking text file can be a damaged analytical object.

The successful part of this checkpoint is a 106-record extraction pilot, with location identities, uncertainty constructs, separator classes and changing handwriting annotations preserved. The unsuccessful part is also explicit: exact-file downloads failed, so the complete upstream transcription was not recovered and its recorded checksum was not verified. We used labelled browser-transferred excerpts instead. Neither that access limitation nor the successful pilot is evidence for a language or decipherment.

1. Continue the investigation without pretending the obstacle disappeared

Volume 2 identified the next task: recover the exact transcription files, verify the relevant source identity, and validate extraction before measuring boundary effects. Its matrix audit showed why the sample itself needs attention. We continue from that checkpoint rather than opening a new interpretive branch. [1]

This run attempted to retrieve ZL3b-n and IT2a-n directly into the working environment. Those requests failed at network name resolution. A separate file-download attempt for ZL also failed. The public source remained readable through the browsing interface. Those outcomes are different and are recorded separately.

We therefore could inspect the transcription, but could not honestly say that we possessed its original downloaded bytes. A text display may preserve what we need for a small local example while differing from the original in whitespace, line endings or omitted material. It cannot simply inherit the original file’s checksum.

The choice was not between declaring success and doing nothing. We could make bounded progress on the extraction procedure, using an explicitly selected sample whose limitations were visible. That is what this volume reports.

The acquisition requirement has not been waived. The complete source-file comparison remains necessary before an exact reconstruction of the older computational input can be claimed. What we have now added is a tested pilot that makes several damaging extraction errors observable before they reach a boundary model.

2. The sample is selected, not representative

The working excerpt contains the displayed locus records from five page units: f1r, f1v, f2r, f2v and f115r. The first four supply a compact range of writing and annotation cases. The last supplies the within-page handwriting changes already flagged by the inherited matrix. This was a purposive software-testing sample, not a random or blinded manuscript sample. [2, 4, 5]

We manually transferred the selected page headers and records from the readable ZL display. Standalone modern commentary was omitted. The local file is named ZL3b-selected-browser-excerpt.txt so that it cannot be mistaken for the complete ZL3b-n source. The transfer history and source locations are retained with the pilot.

These are counts from the local excerpt after extraction. They agree with the corresponding page-level counts retained in Volume 2’s matrix. There were no discrepancies in the checked P, L, C and R totals or paragraph-start totals for these five page units. [4, 5]

That agreement is a useful consistency check. It is not independent validation: the matrix and the displayed records belong to the same upstream transcription family. Nor does agreement in counts establish that every manually transferred character is correct. A substitution can preserve a count while changing the text.

The sample contains no circular- or radial-text loci. It cannot validate the treatment of those populations. The programme still needs broader format coverage and complete-source reconciliation. Our report starts with that limitation so that the smaller achievement can stand without exaggeration.

3. A boundary is not disposable formatting

The format documentation distinguishes confident gaps, doubtful gaps, drawing interruptions, paragraph markers, comments and metadata tags. It also distinguishes the location of a record from its textual payload. These distinctions informed the pilot; they are not meanings we have deciphered from the manuscript. [3]

The deliberately bad cleaner used a generic rule that deletes angle-bracketed strings. Such a rule treats a comment, a paragraph marker and a drawing interruption as though they were interchangeable decoration. Our diagnostic asked what information that rule destroys in the selected input.

The result was exact within that input: all 37 top-level drawing-interruption markers disappeared. They occur in 28 of the 106 selected locus records. The cleaner also removes other angle-bracketed annotations, but the headline count concerns the drawing boundaries alone. [5]

The distinction between 37 markers and 28 affected records matters. Some records contain several interruptions. Counting only affected records would understate the number of boundary events removed; counting markers as independent manuscripts or independent experiments would overstate the evidence.

Consider the invented illustration ab<->cd.ef. It is not a manuscript quotation. Deleting the bracketed interruption produces abcd.ef. The transformation has not merely shortened the file. It has created the contiguous string abcd where the input represented a break between ab and cd.

A later calculation could then count that joined string, estimate its length, or use its ending to define a neighbouring relationship. The output might look orderly while answering a question about the cleaner’s choices rather than the supplied representation.

This is a demonstrated failure mode of the cleaner we constructed. It is not an allegation that a particular published Voynich study made this error, or evidence that an earlier eduKate result was caused by it. Establishing either claim would require inspecting that study’s actual code and inputs.

4. What the preserving parser keeps

The pilot takes a different approach. It records events without discarding their original spelling or their position inside the local payload. A gap becomes a gap event. An interruption remains an interruption. An unresolved alternative remains unresolved. Ordinary text is retained without assigning it a pronunciation or meaning.

Every accepted locus retains its page identity, numbered locus identity, supplied location code, local source-line reference and untouched payload. The event representation records start and end positions, so an investigator can return from an event to the exact substring in the local excerpt.

This makes the preservation check concrete. Concatenating the event strings reproduced the original local payload in 106 of 106 accepted records. A separate reconstruction reproduced the complete local excerpt, including its retained line endings and provenance comments. Direct substring checks also confirmed that the saved event positions point back to their saved strings. [5]

Those are different tests from checking whether the words look plausible. A result can pass a readability check after losing information. A round-trip test instead asks whether the declared transformation can return the representation that entered it.

There is an important limit. We have verified a return to the local transferred excerpt, not to the original online file, the manuscript image or the physical object. A perfect round trip can preserve an imperfect transfer. The complete evidence route therefore still requires a source-byte check and, where a conclusion depends on a doubtful reading, an appropriate return to the image.

The pilot is valuable because it preserves that route rather than claiming to replace it.

5. One page, two recorded handwriting assignments

The handwriting issue at f115r provides a second concrete test. Volume 2 retained the composite value 2+3 instead of forcing the page into one handwriting class. The displayed ZL records show where its annotations change: a tag at the first locus sets H to 2, and a tag at locus 13 sets H to 3. [1, 2]

The format’s metadata rule makes a text tag effective from the locus on which it is set until it is changed or the page ends. The pilot reads those instructions as metadata, not as manuscript glyphs. [3]

Applied to the selected f115r records, that produces 12 loci assigned hand 2 and 33 assigned hand 3. These totals match the inherited matrix’s hand profile. No new identification of a scribe has been made. We have reconstructed the behaviour of the supplied annotations at locus level. [4, 5]

This is a useful distinction for a later controlled analysis. A page-level field saying 2+3 warns that the page is mixed. The locus-level annotation says which supplied assignment applies where. Neither is improved by guessing a single hand for the entire page.

The deliberately bad cleaner would delete the tags. A second kind of faulty implementation could retain them but apply a change only to the next record, or carry the previous page’s final state into a new page. We tested these behaviours with constructed examples. The preserving implementation applies the current record’s tag and resets page-scoped state at the next page header. [5]

Our result concerns the consistency of metadata handling. It does not independently confirm the palaeographic judgement behind the assignments. A competent extraction programme should preserve a scholar’s recorded classification while keeping its evidential origin visible.

6. A text location is not automatically a new physical line

Four f1r records carry a same-line locator and the Pt subtype. The format describes Pt as a right-justified item sharing the preceding item’s physical line. Treating every new record as a new line would lose this relationship. [2, 3]

Our pilot retained all four same-line locators. It did not promote them into four new physical-line starts. It also did not infer sentences from the number of records. [5]

This matters because the inherited research has questions about line boundaries. Before testing what changes at a line start, we need a declared definition of a line start. Counting file rows is not a substitute for consulting the supplied location information.

The distinction becomes clearer with a hypothetical layout. Imagine a paragraph line followed by a short right-aligned item on that same line. A transcription may assign separate location records to the two pieces so that each remains addressable. A model that calls both records independent line openings will be evaluating a boundary partly introduced by its data conversion.

There are several possible analytical choices after preservation. One analysis might keep the two locations distinct. Another might construct a carefully justified shared-line view. A third might exclude ambiguous cases from its primary comparison. What it should not do is make the choice invisibly and describe its output as a direct observation of the page.

This volume implements preservation, not a final historical reading-order model. The later boundary protocol must still decide which relationships it will compare.

7. Sixty-five doubtful seams change the diagnostic count

The selected input contains 65 top-level doubtful-gap events. We generated two explicitly labelled analytical views from the same preserved events. One treats each doubtful seam as a split; the other joins it. Both retain confident gaps and drawing interruptions as separators, and neither joins separate locus records. [5]

The split view contains 934 diagnostic spans. The joined view contains 869. The difference is exactly 65, matching the number of eligible doubtful seams in this sample. [5]

These are not two rival estimates of how many linguistic words Voynich contains. They are two outputs of stated segmentation policies on a selected excerpt. Bracketed alternatives and ligature notation remain opaque in this diagnostic; no alternative reading is selected to improve a result.

The calculation gives a simple check on the implementation. Each doubtful seam in this accepted sample lies between text-bearing material. Switching its treatment therefore adds or removes one diagnostic boundary. If the difference had not matched the eligible seam count, the discrepancy would have required investigation before using either view.

It also gives a useful research warning. A boundary policy changes the population of spans on which later statistics operate. Comparing two headline results without accounting for that population change can conceal a change in the underlying question.

Suppose a later model performs differently under the two views. We would first ask where their eligible events differ, which exact source locations drive the difference, and whether the difference survives a predefined comparison. We would not immediately declare that one view has revealed the true language.

The present result does not decide whether the doubtful seams are meaningful historical boundaries. It shows that the implementation preserves the choice and makes its counting consequence inspectable.

8. Uncertainty survives extraction

The top-level event census includes ten alternative-reading constructs, nine ligature constructs, six uncertain-mark events and two standalone extended-code events. These figures describe the pilot’s representation, not the complete inventory of unusual glyphs. [5]

The qualification “top-level” is essential. An extended code inside a bracketed alternative remains inside that opaque alternative. It is not counted again as a standalone event. Without that qualification, a reader could mistake a parser category for an exhaustive count of every embedded mark.

The programme does not choose the first member of an alternative, flatten every connected form into letters, or delete an uncertain character because a downstream model expects a clean string. Those operations may be proposed as separately recorded analytical views, but they are not allowed to replace the preserved input unnoticed.

There is a trade-off. An opaque uncertainty construct is inconvenient for some calculations. Keeping it may reduce the number of immediately eligible examples. That is preferable to gaining a larger sample by silently converting uncertainty into certainty.

For the future q/y work, this means that an uncertain ending or prefix needs its own eligibility rule. The researcher should be able to identify precisely which records were excluded, expanded into alternatives or retained under a particular interpretation. The present pilot does not define the final inclusion rule or report a q/y effect.

What it supplies is the infrastructure needed to make that rule accountable.

9. Refusal is part of the extraction result

This implementation is deliberately a bounded merged-EVA pilot, not a complete IVTFF parser. It accepts a declared input profile and refuses several structures it does not yet support. Wrapped records and interlinear alignment are among the unimplemented cases.

That refusal is a limitation, but it is also a useful safety property. A programme that silently skips an unfamiliar line can return a plausible sample while concealing what disappeared. A programme that stops with an identifiable reason makes the missing capability visible.

The checks also reject duplicate page or locus identities, out-of-order locus numbering within a selected page, conflicting metadata changes, malformed brackets and several inconsistent paragraph-marker arrangements. These are implementation rules tested in this pilot. They do not prove that every syntactically accepted source record is historically correct. [5]

No file-wide claim is made about every permissible IVTFF construct. Some valid input may be outside this support profile. The next development step must expand the profile deliberately, with corresponding examples and tests, rather than weakening the refusal rules simply to obtain a larger accepted count.

This distinction prevents another form of overstatement: “the parser ran” is not the same as “the entire intended corpus was correctly interpreted”. Full-corpus acceptance needs an accounting of every accepted record, every rejected record and every explicitly unsupported construction.

10. What the 32 software tests actually establish

The pilot completed 32 constructed software tests. They exercise local round trips, interruption preservation, comment handling, bracketed alternatives, metadata persistence and reset, identity checks, paragraph-state checks, unsupported-input refusal and the two doubtful-seam policies. All passed in the recorded run. [5]

A separate small counting path checked the selected number of loci, drawing markers and paragraph markers directly. It agreed on 106 records, 37 drawing markers, 23 paragraph starts and 23 paragraph ends. The additional position checks confirmed the stored event offsets against the local payloads. [5]

These are checks of selected implementation behaviours. They are not 32 independent experiments establishing a manuscript hypothesis, 32 expert endorsements, or a measured probability that the programme contains no bugs. Both code paths were prepared within this investigation; no independent external reviewer is claimed.

The sample was also used during development. It is not an untouched held-out evaluation set. A later test on new material should preserve that distinction rather than counting the development examples again as fresh confirmation.

The most defensible statement is therefore narrow: the implementation passed the recorded checks, preserved the tested local representations and exposed the deliberately introduced destructive-cleaning failure. Wider correctness remains a task for broader validation.

11. Why this is progress rather than another discussion of representation

The existing research already explains that transcription choices matter. This volume does not claim to have discovered that general principle. Its contribution is an executed comparison on a retained, identified local sample.

The difference is practical. We now have a case where one specified cleaning operation removes 37 recorded boundaries. We have a preserving event representation whose local reconstruction can be checked. We have recovered the supplied locus-level transition between two handwriting assignments on f115r. We have quantified the effect of a doubtful-seam policy on the diagnostic sample. [5]

Each result points to material that can be inspected again. A reader can ask which interruption disappeared, which record received a handwriting assignment, or why a span exists in one view and not another. The answer no longer has to be an attractive explanation about “structure”. It can be a source address, an event and an explicit transformation.

None of this establishes the historical generating mechanism. A sound extraction can feed a mistaken theory. But an unsound extraction can undermine even a promising theory before its predictions are fairly tested.

Our pathway therefore separates two duties: preserve what the input records, then challenge the interpretation applied to it. The first duty does not solve the second. It makes the second more testable.

12. What remains open after this volume

The full source-byte comparison remains incomplete. The expected ZL digest is retained from the earlier manifest; there is no verified complete downloaded file against which to compare it. The failed access attempts have not been reclassified as either a source mismatch or evidence of a changed transcription.

The original event tables and calculation code behind the older boundary odds ratios have not been recovered here. No current-token final-y effect, preceding-token effect or oR/qoR association has been estimated in this pilot. We have not converted an extraction check into a scientific replication.

Test 2A, the blind recurring-component investigation, also remains separate and unfinished in this record. No visual components were coded, no manuscript-image classifications were independently reassessed, and no text–image semantic relationship was tested.

The point of this table is to prevent a successful component from lending its status to everything around it. The extraction pilot is complete within its stated scope. The acquisition task is not. Both facts belong in the same dated record.

13. The next experiment now has a better entrance

The next step still requires a byte-preserving acquisition of the intended corpus, or an explicitly declared new input edition when the intended source cannot be recovered. An available mirror could be a candidate, but its ancestry and bytes would need checking; it should not automatically inherit the identity of the original URL.

Once a complete candidate is available, the programme should retain the original bytes before any normalisation, compare the expected digest, and record any disagreement. A different line-ending convention may explain a mismatch, but the explanation must be investigated rather than assumed. The untouched candidate and any normalised derivative need separate identities.

The parser should then be evaluated on the entire intended input. Unsupported constructions must be addressed explicitly. Accepted locations and page-level totals should be reconciled with the inherited accounting layer, with every difference recorded. An apparent count match should not substitute for checks of location identities, payload preservation and metadata scope.

Only after that should the detailed boundary experiment freeze its eligibility rules, comparison conditions and evaluation split. The current pilot should be treated as development material. Its successful examples should not be presented as an unseen scientific test set.

This may sound slower than producing a new interpretation immediately. It is more direct in a different sense: it removes a specific way that our own processing could manufacture the pattern we later become excited about.

We have not made the manuscript speak. We have made one part of the research procedure show what it would otherwise erase.

That is the new step carried into the catalogue.

Sources and reproducibility record

[1] Published predecessor. Voynich | edkSG Vol. 2 — The Input Audit, EDKSG-VOY-V002, edition 1.0.0. The publication identity was checked before continuation.

[2] Primary transcription display. Zandbergen–Landini ZL3b-n, displayed header identifying version 3b, 13 May 2025. The selected records are f1r.1–28, f1v.1–10, f2r.1–15, f2v.1–8 and f115r.1–45, with their page headers. Browser text inspected 5 September 2026; complete raw-file identity was not verified.

[3] Format authority. René Zandbergen, IVTFF specification, particularly comment handling, Table 9 and sections 6.7–6.9. Relevant page images were inspected. The pilot is a bounded implementation informed by the specification, not a claim of complete format conformance.

[4] Inherited comparison data. Volume 2’s retained matrix-rows.json and run manifest, derived from the frozen Voynich_Test_1b_Frozen_Segmentation_Matrix.xlsx. These supply the per-page comparison counts and the inherited f115r hand profile; they are not independent source observations.

[5] This volume’s executed record. EDKSG-VOY-V003-PILOT-01. The accompanying research package contains the selected browser-text transfer, source-selection manifest, extraction_gate.py, parsed records, pilot results, 32-test run log, narrow second-path checks and acquisition-failure record. The data-processing results above refer to that run. The complete third-party transcription corpus and manuscript images are not included.

Further orientation. EVA, Transcription and the Segmentation Problem remains the general introduction. Volume 1 — The Research Baseline preserves the start of this chronological series.

Catalogue: Voynich Research Library · Editorial framework: Wintour House.

EDKSG-VOY-V003 · Edition 1.0.0 · 5 September 2026 · Predecessor EDKSG-VOY-V002 · Baseline EDKSG-VOY-B000. Earlier volumes remain unchanged. Corrections must identify the affected result and edition. No independent scientific replication or decipherment is claimed.