Voynich | The Physical Manuscript and Its Digital Double

VOYNICH · CUSTODY THROUGH TIME · MOVEMENT II

The physical Voynich Manuscript is one object. The digital Voynich is a family of derived objects that can travel farther, multiply faster and sometimes reveal more—while never becoming the parchment they represent.

Research thesis: Modern Voynich scholarship routinely speaks as though everyone were studying the same manuscript. In one sense, we are. In another, we are not. Beinecke MS 408 is a single material codex in New Haven. What most researchers actually touch are representations: institutional images, facsimiles, screenshots, crops, transcriptions, normalised corpora, feature tables and machine-readable derivatives. Each is a transformation with its own strengths, losses and lineage. Serious research must therefore identify not only the object being studied, but the representation through which the object became measurable.

This opens Movement II of Voynich | Custody Through Time — The River, the Handoff and the Future Receiver. It follows The Manuscript That Almost Stopped Moving, where the material river slowed while the representation river accelerated.

One Manuscript, Many Research Objects

The physical Voynich Manuscript is not a category.

It is one particular object.

V_M = BEINECKE MS 408

one parchment codex
one material history
one current institutional coordinate

The digital Voynich is different.

There is no single digital object that exhausts it.

V_R = {R₁, R₂, R₃ ... Rₙ}

where each \(R_i\) is a representation derived directly or indirectly from the material codex.

an institutional high-resolution image;
a resized web image;
a screenshot;
a rotated folio;
a cropped glyph;
a colour-adjusted image;
a printed facsimile;
an EVA transcription;
an alternative transcription;
a normalised token stream;
a statistical feature matrix;
a graph derived from co-occurrence;
a machine-learning input;
an AI-generated description.

All can truthfully be described as being “about” the Voynich Manuscript.

They are not epistemically identical.

We share one source object. We do not always share one research representation.

The Digital Double Is Not a Clone

The phrase digital double can mislead if it suggests perfect duplication.

A digital image does not reproduce the manuscript’s animal skin, weight, odour, flexibility, thickness, pressure relief, microscopic deposition, true translucency, binding mechanics or three-dimensional surface behaviour.

What it reproduces is a selected set of measurable visual signals under declared or undeclared imaging conditions.

OBJECT
→ illumination
→ optics
→ sensor
→ colour processing
→ image file
→ display
→ observer

The double is therefore better understood as an instrument-produced state of the object.

It can be extremely faithful to the dimensions it was designed to preserve.

It can also be silent about dimensions it never measured.

A Representation Is a Measurement

This is the first methodological correction of Movement II.

We often talk about a scan as if it were merely a transparent window.

It is better treated as a measurement produced by an imaging system.

R = T(V_M; I)

where \(V_M\) is the material manuscript, \(I\) is the imaging configuration, and \(T\) is the transformation that produces representation \(R\).

The configuration can include:

illumination spectrum;
illumination angle;
camera geometry;
lens characteristics;
sensor response;
focus;
exposure;
colour calibration;
resolution;
file compression;
post-processing;
cropping and framing.

A result derived from \(R\) therefore inherits the properties of \(T\) whenever those properties touch the measured feature.

The Original Is Not Automatically Better for Every Question

There is an equally important correction in the opposite direction.

“Original” should not become a prestige label that automatically defeats representation-based research.

A digital image may make certain visual comparisons easier and more reproducible than a brief physical consultation.

Magnification can make minute visible forms easier to compare.
Two folios can be placed side by side instantly.
Coordinates can be measured repeatedly.
Image-processing experiments can be versioned.
Thousands of page locations can be sampled computationally.
Multiple observers can receive the identical image file.

The question is therefore not:

Which is more authentic?

The scientific question is:

Which representation preserves the signal required by this hypothesis?

The Object Is the Anchor Datum

The fact that a scan can outperform casual physical viewing for some tasks does not remove the special status of the original.

The physical manuscript remains the anchor because the derivative chain points back to it.

V_M
↓
R₁
↓
R₂
↓
R₃
↓
MEASUREMENT

If \(R_3\) becomes questionable, researchers can inspect \(R_2\).

If \(R_2\) contains a transcription ambiguity, they can return to \(R_1\).

If the image itself cannot discriminate the physical alternatives, the chain can—in appropriate circumstances—return to \(V_M\).

The original is not valuable because every question requires touching it. It is valuable because every derivative ultimately needs somewhere real to return.

The Representation Graph

The digital Voynich should not be modelled as one straight line.

It branches.

                 → crop → glyph dataset → classifier
                /
V_M → master image → web image → screenshot → article figure
                \
                 → transcription A → normalised corpus → statistics
                  \
                   → transcription B → alternate segmentation → statistics

Different branches can produce different downstream answers while remaining descendants of the same material object.

This is not necessarily error.

It may be the correct expression of different measurement questions.

But the branches must remain identifiable.

Representation Provenance

Physical provenance asks:

Where did the manuscript come from?

Representation provenance asks:

Where did this particular digital or analytical state come from?

For a serious derivative, record where possible:

source object;
source image identifier or URL;
imaging institution;
image version or date;
download date if relevant;
crop coordinates;
rotation;
rescaling;
colour or contrast transformation;
transcription regime;
segmentation policy;
normalisation rules;
feature extraction method;
code version;
downstream dataset version.

Without that chain, two researchers can appear to disagree about Voynich while actually measuring two different derivative states.

Representation Distance

Movement II needs a way to discuss how many transformations separate a claim from the material datum.

Call this representation distance.

D_R = number and severity of transformations
between V_M and the measured state

A direct institutional image has lower representation distance than a screenshot of a compressed image embedded in a secondary article.

A raw transcription has lower distance from visible marks than a linguistic feature table derived after token normalisation.

But distance alone does not determine quality.

A carefully documented higher-distance derivative can be scientifically excellent for the job it performs.

An undocumented low-distance crop can be misleading.

Representation distance is not a guilt score. It is a reminder to carry the transformation history.

Fidelity Is a Vector, Not One Number

A representation can be highly faithful in one dimension and poor in another.

F(R) = (
geometry,
colour,
texture,
scale,
sequence,
materiality,
context,
transcription ambiguity,
metadata
)

A high-resolution scan may have excellent visible geometry and page context, strong colour under calibrated conditions, but almost no direct information about pressure relief or page flex.

A facsimile may restore scale, page turning and foldout experience while remaining weak for microscopic pigment questions.

An EVA transcription is highly useful for repeatable text statistics while deliberately discarding much of the visual stroke state.

Therefore the proper question is not “How faithful is this representation?”

It is:

Faithful to which property?

The Representation Affordance Matrix

Different representations enable different research operations.

Representation	Strong affordances	Important blind spots
Original manuscript	Material construction, surface, pressure, true scale, translucency, stitching, microscopic evidence, intervention sequence	Limited access; cannot be distributed or handled at scale
Institutional high-resolution image	Visible geometry, layout, colour relationships, magnification, repeatable page comparison, global access	Physical relief, some material layering, page flex, true tactile relationships
Facsimile	Approximate size, codex sequence, foldout experience, page relationships	Original materials and microscopic surface evidence
Crop	Focused comparison, blinded scoring, machine vision	Page context, scale and neighbouring relationships may disappear
Transcription	Search, counts, sequence statistics, reproducibility	Visual ambiguity and transcription choices become encoded assumptions
Normalised corpus	Large-scale modelling, cleaner statistical comparison	Variant forms and uncertainty may be removed
Feature table	Efficient quantitative analysis	Researcher sees features rather than manuscript state
AI summary	Fast synthesis and navigation	Can detach conclusions from claim state, source lineage and uncertainty

There is no universally best row.

The scientific job selects the representation.

SCAN-TESTABLE Is Not the Same as Image-Only Forever

The custody pillar introduced:

TESTABILITY = {SCAN, OBJECT, BOTH}

Movement II sharpens what that means.

A question is SCAN-TESTABLE when the relevant signal is preserved sufficiently in the declared digital representation.

That does not mean every possible scan is adequate.

Resolution, spectral range, colour calibration and framing can matter.

A question is OBJECT-TESTABLE when the decisive discriminator resides in material evidence not adequately represented.

A question is BOTH when digital analysis narrows the hypothesis but material inspection may be required for final discrimination.

The category belongs to the question, not permanently to the topic.

A Digital Image Can Create New Observables

Representation does not only remove information.

It can create new analytical affordances that do not exist in the same form during ordinary physical viewing.

Pixel coordinates create a stable measurement grid.
Zoom creates repeatable magnification.
Image alignment can compare distant folios.
Digital masking can blind semantic cues.
Automated extraction can sample thousands of repeated locations.
Image versions can be hashed and shared.

This means the transformation:

OBJECT → DIGITAL IMAGE

is not simply:

RICH → POOR

It is:

ONE AFFORDANCE SET
→
A DIFFERENT AFFORDANCE SET

Some dimensions are lost.

Some become easier to measure.

The Crop Problem

A crop is one of the smallest and most dangerous transformations in visual Voynich research.

It seems innocent.

Yet cropping decides what the receiver is allowed to see.

FULL PAGE
→ SELECTED REGION
→ RECEIVER ATTENTION

A crop can remove:

folio identity;
scale;
neighbouring text;
repeated motifs;
page position;
damage context;
section identity;
visual cues that would weaken the proposed interpretation.

This is why blinded visual work should preserve a reversible map from crop back to source folio even when scorers are intentionally denied that map during scoring.

Blinding removes information from the observer.

Good lineage preserves it for the experiment.

The Transcription Is a Second Representation Revolution

An image remains recognisably page-like.

A transcription changes the ontology of the object.

VISIBLE MARKS
→ DISCRETE SYMBOLS
→ MACHINE-COUNTABLE SEQUENCE

This is enormously productive.

It is also where segmentation assumptions become structurally powerful.

The EVA Transcription and the Segmentation Problem already treats this explicitly. Movement II now places that problem inside the wider representation chain.

A transcription does not discover the units by merely typing them.

It proposes a repeatable representation of the marks.

The transcription is not Voynichese. It is our instrument for making Voynichese countable.

Normalisation Can Remove the Very Signal We Need

Data cleaning is usually treated as improvement.

Sometimes it is.

But an unknown script creates a special danger.

If two visible variants are normalised into one symbol before we know whether the distinction matters, the dataset can erase a potentially meaningful contrast.

VARIANT A + VARIANT B
→ NORMALISED X
→ DIFFERENCE LOST

The same applies to uncertain spaces, damaged forms, overwritten signs and ambiguous strokes.

Cleaning should therefore be reversible whenever possible.

The Derived Dataset Can Become More Famous Than Its Source State

Once a convenient corpus exists, downstream researchers may stop returning to the images.

This is efficient.

It can also cause representation assumptions to become invisible.

IMAGE CHOICE
→ TRANSCRIPTION CHOICE
→ SEGMENTATION CHOICE
→ NORMALISATION CHOICE
→ CORPUS
→ 100 PAPERS

By the hundredth paper, the first four choices may no longer be discussed.

The dataset begins to feel like the manuscript.

This is representation normalcy blindness.

The correction is not to abandon shared corpora.

It is to preserve their ancestry.

Digital Derivatives Need Custody Too

The physical manuscript has an institutional custodian.

Derived states are often much more distributed.

V_M
→ one primary institutional custody

V_R
→ many distributed custodians

A repository may host a transcription.

A researcher may maintain a corrected corpus.

A paper may freeze one version.

An archive may preserve another.

An AI system may ingest an unlabelled copy.

Once representation custody becomes distributed, version identity matters as much as physical shelfmark identity.

A Shelfmark for Data

Beinecke MS 408 gives the physical codex a stable institutional identity.

Digital research needs analogous discipline.

Not necessarily one universal naming system, but enough persistent identity to answer:

Which image set?
Which transcription?
Which correction state?
Which normalisation?
Which code?
Which derived feature set?
Which frozen test?

The principle is simple:

A result that cannot identify its representation cannot be fully reproduced.

Representation Hashes and Frozen Tests

Digital objects permit a preservation technique unavailable to historical owners: exact file or content hashes.

A hash cannot prove that the representation faithfully captures the manuscript.

But it can prove that two analyses used the same frozen digital payload.

OBJECT FIDELITY
requires calibration

DIGITAL IDENTITY
can be frozen cryptographically

This distinction has already mattered in the eduKate Voynich experiments. A corrected payload creates a new versioned challenge. It does not retroactively repair a failed frozen test.

That is digital custody discipline.

The Digital Double Can Fork

Physical Beinecke MS 408 cannot be casually duplicated into competing originals.

Digital descendants can fork instantly.

R₀
├── R₁ corrected transcription
├── R₂ alternate segmentation
├── R₃ colour-adjusted image set
└── R₄ machine-readable normalisation

Forking is often desirable.

It allows competing models to coexist.

The danger begins when lineage disappears and one branch quietly presents itself as the source.

The Digital Double Can Also Outlive Its Explanation

A file may survive longer than the web page describing how it was made.

A dataset may survive after its README disappears.

A model may retain numerical features after the transcription policy that produced them is forgotten.

This is the digital version of the Lost Runtime Problem.

DATA SURVIVES
PROTOCOL DIES

The same civilisational failure that may have affected the original manuscript can therefore recur around its digital double.

Digitisation does not abolish the handoff problem. It gives the handoff problem new substrates.

A Future Human May Receive the Double but Not the Original Context

Imagine a future receiver in 2300 recovering a large bundle of Voynich files.

They have:

images;
transcriptions;
statistics;
old PDFs;
AI embeddings;
archived web pages.

But suppose the links between them are gone.

Which transcription corresponds to which scan?

Which corpus was corrected?

Which statistics were measured before the correction?

Which labels were conventional shorthand rather than semantic conclusions?

The future could inherit enormous quantities of information and still lose knowledge about how the representations relate.

That is why representation lineage is part of the future handoff.

The Digital Double Is Not One Generation

Representations themselves have generations.

GEN 0 — physical manuscript
GEN 1 — direct imaging / direct facsimile
GEN 2 — crop / resize / transcription
GEN 3 — normalised corpus / feature extraction
GEN 4 — model / synthesis / AI representation
GEN 5 — downstream summaries of those outputs

Generation number does not determine truth.

It determines what reconstruction information should accompany the claim.

The farther downstream we go, the more important it becomes to retain a route back upstream.

The Return Path

Every derivative research claim should have a return path.

CLAIM
→ measurement
→ feature set
→ corpus/transcription
→ image
→ folio
→ material object

Not every claim needs to traverse the whole route during ordinary use.

But the route should remain reconstructable when the claim is challenged.

This is the representation equivalent of provenance.

When the Return Path Breaks

A claim becomes fragile when its return path breaks.

STATISTIC
→ ?
→ ?
→ UNKNOWN CORPUS
→ ?
→ "VOYNICH"

The number may still be correct.

But its evidential custody has weakened.

We can no longer tell which object state the number describes.

The Digital Double Changes the Meaning of Reproducibility

For a physical manuscript, exact repeated access is difficult.

The object ages.

Lighting differs.

Instruments change.

Digital representations create stronger reproducibility in another sense.

Two researchers can run the same code against the same frozen file.

SAME REPRESENTATION
+
SAME METHOD
→
REPRODUCIBLE COMPUTATIONAL TEST

But computational reproducibility is not the same as object validity.

REPRODUCIBLE ON R
≠
R CAPTURES THE REQUIRED PROPERTY OF V_M

Both tests are required.

The Digital Double Can Be More Precise Than the Historical Question

There is another form of representation arrogance.

A modern image or dataset can be extremely precise.

The historical question may still be low-resolution.

We can measure a glyph coordinate to the pixel and still not know whether the original maker regarded that stroke as a separate unit.

We can estimate a colour value precisely and still not know the historical name, mixture or intended significance of the pigment.

Measurement precision should not inflate historical resolution.

A high-resolution representation does not automatically create a high-resolution historical meaning.

The Double Can Become the Default Reality

Most contemporary researchers will see Beinecke MS 408 first—or only—through a screen.

That makes the digital view normal.

Normality makes representation choices harder to notice.

Page orientation, background colour, cropping, zoom behaviour and interface order begin to feel like properties of the manuscript itself.

This is why the Tangential Voynich work on Normalcy Blindness belongs here too.

The most influential representation is often the one we stop recognising as a representation.

The Physical Original Protects Us From Representation Closure

If only derivative files remained, later researchers might be forced to accept the measurement choices encoded into those files.

The continuing existence of the original prevents total closure.

A future imaging method can ask the object a question none of our present images were designed to answer.

A future conservator can inspect a material relationship our transcriptions never recorded.

A future scholar can challenge our segmentation by returning to the marks.

The original keeps future questions possible.

Object Preservation Is Preservation of Unasked Questions

This is perhaps the deepest conservation argument of Movement II.

We preserve the original not only because it contains answers.

We preserve it because future generations may have better questions.

PRESERVE OBJECT
=
PRESERVE PRESENT EVIDENCE
+
PRESERVE FUTURE OBSERVABILITY

A scan freezes one interrogation of the material state.

The object preserves the possibility of another.

The Digital Double Protects the Original Too

The relationship is therefore reciprocal.

The original anchors the digital double.

The digital double protects the original by absorbing enormous quantities of routine research demand.

ORIGINAL
anchors
REPRESENTATION

REPRESENTATION
shields
ORIGINAL

This is one of the great achievements of modern manuscript custody.

The object and representation are not rivals.

They form a preservation-research pair.

A Researcher Must Declare Which Voynich They Used

From Movement II onward, a serious Voynich result should be able to answer:

What is the source object?
What exact representation was used?
How was that representation produced?
What transformations were applied?
What properties does the representation preserve well?
What relevant properties may it flatten or omit?
Can the transformation be reversed or reconstructed?
Can the analysis return to a higher-fidelity source if challenged?

This is the representation ledger.

The Representation Ledger

For experiments where representation choice can affect the result, record:

REPRESENTATION ID
SOURCE OBJECT
SOURCE VERSION
TRANSFORMATION CHAIN
RESOLUTION
COLOUR / SPECTRAL STATE
CROP / ROTATION
TRANSCRIPTION REGIME
SEGMENTATION REGIME
NORMALISATION
KNOWN LOSSES
EXPECTED AFFORDANCES
RETURN PATH

This joins the Observer Ledger, Null Ledger and Defeat Ledger already developed in Tangential Voynich.

OBJECT LEDGER
+
REPRESENTATION LEDGER
+
OBSERVER LEDGER
+
NULL LEDGER
+
DEFEAT LEDGER

The purpose is not bureaucracy.

It is to stop invisible transformations from masquerading as properties of the manuscript.

What This Changes for Padua

Padua hypotheses now need a representation declaration too.

If a claim concerns visible page geometry, the Yale images may be sufficient.

If a claim concerns subtle ink sequence, pressure, parchment preparation or microscopic material behaviour, a digital image may be only the discovery layer.

If a localising argument depends on colour, the colour representation and calibration state become part of the evidence.

If it depends on transcription statistics, the transcription version and segmentation assumptions become part of the provenance claim.

A hypothesis cannot be more object-specific than the representation that supplied its evidence.

What This Changes for AI

AI usually receives neither parchment nor even institutional master images directly.

It receives derivatives.

OBJECT
→ IMAGE
→ CAPTION
→ ARTICLE
→ DATASET
→ TRAINING CORPUS
→ MODEL REPRESENTATION

By the time the model answers a question, several representation layers may already have disappeared from view.

The Voynich AI Ingestion Guide therefore becomes part of custody-through-time: source claim → ingested claim must never increase certainty, and representation lineage should remain available whenever it materially affects the claim.

The Future Receiver Will Inherit Both

The future does not receive only Beinecke MS 408.

It also receives our representations of Beinecke MS 408.

Those representations may become historically important in their own right.

A future scholar may study a 2004 Yale image set as evidence of how the manuscript was imaged and perceived in the early digital era.

They may study EVA as evidence of how twenty-first-century scholars chose to carve the visible script into machine-readable units.

They may study our AI datasets as evidence of which distinctions our generation considered important enough to encode.

The double eventually becomes part of history too.

The Double Is Another Handoff

A digital representation is not merely an access product.

It is a handoff packet from one custodial state to another receiver.

MATERIAL CUSTODIAN
→ REPRESENTATION
→ REMOTE RECEIVER

The packet succeeds when the receiver can recover the signal required for the intended task without being misled about what the packet does not contain.

That is the Receiver Completion Law applied to digitisation.

A High-Fidelity Digital Handoff

A high-fidelity digital handoff should carry more than pixels.

object identity;
folio identity;
imaging provenance;
technical metadata where relevant;
orientation;
scale or dimensions;
known limitations;
rights and custody information;
links back to the institutional source;
version information.

The better this packet is, the less likely a derivative researcher is to confuse the representation with a complete material substitute.

The Digital Double Does Not Reduce the Mystery

It changes which parts of the mystery can be attacked cheaply and repeatedly.

It lets us count, compare, crop, blind, align, search and model at scales unavailable to earlier custodians.

But if the missing discriminator lives in a material property, no amount of repeating the wrong representation will create it.

MORE ANALYSIS OF R
≠
NEW INFORMATION ABSENT FROM R

That is why representation limits must be declared before interpretation becomes confident.

Movement II’s Central Question

Movement I asked where the object is and how it arrived there.

Movement II asks:

When we say “the Voynich Manuscript,” which layer of the Voynich Manuscript are we actually holding?

Sometimes the answer is:

the material codex.

More often today, it is:

a representation
of a representation
of a measurement
of the codex

That is not a reason to distrust modern research.

It is a reason to make the route explicit.

World Return

The physical Voynich Manuscript remains one object in New Haven.

Its digital double is not one object.

It is a branching family of representations.

Some preserve visible geometry beautifully.

Some preserve codex experience.

Some preserve sequences for computation.

Some preserve only selected features.

Each enables questions.

Each closes others.

The original remains because future questions may require signals our representations did not know to preserve.

The representations spread because present questions can often be answered without spending the original.

The original preserves the possibility of new observation. The digital double preserves the possibility of wide observation. Modern Voynich research needs both—and must never forget which one supplied the evidence.

The next article follows the consequence outward:

Voynich | When the Object Stayed and the Information Escaped →

Research routes: Custody Through Time · The Manuscript That Almost Stopped Moving · Where Is the Manuscript Today? · Voynich Research Library