eduKateSG · VOYNICH · CUSTODY THROUGH TIME · MOVEMENT II
Voynich | What Can a Scan Know?
A scan can preserve extraordinary evidence. It can also make us forget that we are looking through an instrument.
← When the Object Stayed and the Information Escaped · The Physical Manuscript and Its Digital Double · Voynich Research Library
Most people who study the Voynich Manuscript today do not study the physical manuscript.
They study images of it.
That statement is not a criticism. It is one of the great achievements of modern conservation. Yale’s Beinecke Rare Book & Manuscript Library provides high-resolution images of the complete manuscript so that researchers can inspect the visible codex without repeatedly handling the fragile original. Yale explicitly encourages scholars to consult the scans and existing research before requesting original-object access.
But this success creates a methodological trap.
The better the digital image becomes, the easier it is to forget that an image is a measurement made under conditions.
A scan is not the manuscript. A scan is an observation channel from the manuscript.
That distinction changes almost every serious Voynich claim.
If we measure glyph geometry, page layout or visible recurrence, a high-quality scan may carry nearly all the signal we need.
If we ask about microscopic ink deposition, parchment surface, pressure, translucency, sewing mechanics, pigment layering or a subtle erased intervention, the public image may carry only part of the relevant state—or none of it.
The central question of this article is therefore deliberately narrow:
What claims can a scan legitimately support before we begin pretending the representation knows what only the object can know?
Yale’s current policy and image access: The Beinecke Cipher (Voynich) Manuscript.
The First Rule: Pixels Are Measurements, Not Parchment
A digital manuscript image begins with a physical event.
Light interacts with parchment, ink, pigment, dirt, abrasion, folds and surface geometry. An optical system collects some of that reflected or transmitted signal. A sensor converts incoming light into electrical measurements. Processing then maps those measurements into pixels, colour values and files suitable for storage and display.
We can write the chain schematically:
OBJECT → illumination → optics → sensor → calibration → processing → file → viewer → display → observer
Every arrow has conditions.
- illumination spectrum;
- light direction;
- camera angle;
- focus;
- lens distortion;
- sensor resolution;
- dynamic range;
- colour calibration;
- white balance;
- image sharpening;
- compression;
- downsampling;
- viewer rendering;
- screen calibration;
- zoom level.
A pixel value is therefore not a tiny piece of parchment stored inside a computer.
It is the output of a measurement and transformation system.
pixel ≠ material point
This does not make pixels unreliable. It tells us what kind of reliability they can have.
A well-calibrated image can be an extraordinarily precise instrument for visible form. The scientific discipline is simply to remember which physical variables were projected into that image and which were not.
A Scan Knows What Survived the Imaging Channel
Suppose the full material state of a page is represented by O.
The scan is not O.
It is a transformation:
S = I(O | C)
where I is the imaging process and C is the set of imaging conditions.
A claim is strongly scan-testable when the signal needed to evaluate it survives that transformation with enough fidelity.
This produces a simple but powerful rule:
SCAN-TESTABLE does not mean “visible somehow.” It means “the representation carries the decisive signal with adequate resolution and known enough conditions.”
The same manuscript question can move between testability classes as imaging improves.
A faint mark invisible in an ordinary photograph may become visible in another spectral band. Surface relief poorly captured under flat lighting may become measurable with raking light or reflectance-based methods. Material composition inaccessible to ordinary photography may be addressed using spectroscopy or X-ray fluorescence.
So “the scan” is not one universal instrument.
Every image set has an instrument profile.
What High-Resolution Images Can Usually Carry Well
For many of the questions at the centre of computational and visual Voynich research, high-quality digital imagery is not a compromise. It is the appropriate research instrument.
Visible glyph geometry
Images can support comparison of visible stroke shape, relative proportions, loop forms, apparent joining, spacing, local variants and recurrent graphic components—subject to resolution, focus and the ambiguity of handwritten forms.
Line and paragraph architecture
Line starts, line ends, paragraph starts, indentation, visible spacing, text blocks, label placement and many boundary relationships are strongly image-accessible.
Page layout
The relation among image fields, labels, prose blocks, radial diagrams, margins and many foldout components can be studied visually at high fidelity.
Visible recurrence
Repeated geometric components, visual motifs and page-level similarities can be compared across the manuscript without repeatedly exposing the original.
Visible damage and interventions
Many tears, holes, stains, missing areas, folds, later numbers, marginal additions, erased-looking zones and contact marks are visible enough to generate testable hypotheses.
Relative spatial position
Coordinates within a page image can be used to ask where glyphs, labels and visual components occur relative to one another—provided page rotation, cropping and scaling are controlled.
Transcription auditing
Scans allow researchers to return from a machine-readable transcription to the visible source and ask whether a symbol, gap or boundary was recorded consistently.
These jobs account for a large part of what modern Voynich analysis does.
For visible structure, the digital representation can often be the primary working instrument even though the physical codex remains the ultimate source object.
But “Visible” Is Not One Variable
Even apparently simple visible claims can depend on image conditions.
Consider colour.
A screen can display red, green and blue values precisely. That does not mean the displayed colour is a direct chemical measurement of pigment. The visible result depends on illumination, sensor response, colour profile, processing and the user’s display.
same-looking colour ≠ same pigment
different-looking colour ≠ necessarily different pigment
Yale’s linked scientific analysis of Voynich inks and pigments illustrates the difference beautifully. Chemical and instrumental investigation can identify material constituents that ordinary colour inspection cannot establish merely by appearance.
The same distinction applies to faint marks.
A mark can be physically present yet absent from one digital representation because contrast is insufficient.
A pattern can appear in a compressed derivative because processing accentuates edges.
A line can appear disconnected because resolution removed a narrow bridge.
A tiny gap can appear meaningful because a viewer enlarged an interpolation artefact.
This is why visible evidence should carry an image-state description whenever the distinction matters.
Yale links the manuscript’s chemical ink and pigment analysis from its main Voynich collection page. See the institutional Voynich access page and its linked analysis report.
The Most Dangerous Error: Scan Absence → Object Absence
Negative evidence is where representation mistakes become especially serious.
A researcher searches an image and does not see a feature.
The sentence then quietly becomes:
the manuscript does not contain the feature.
That inference is licensed only if the image had a fair chance to carry and reveal the feature.
The correct logic is:
not detected in representation R under conditions C
before any stronger conclusion is considered.
To promote scan absence into object absence, we need a detection argument:
- the proposed feature should be visible in this imaging modality;
- its expected size exceeds the effective resolution floor;
- contrast should be adequate;
- the relevant region is in focus;
- the image is not obscured by show-through or surface damage;
- compression has not removed the signal;
- the feature is not spectral, chemical, subsurface or three-dimensional in a way ordinary imaging cannot capture;
- the search procedure itself has adequate sensitivity.
Without those conditions, the honest state is:
NOT OBSERVED IN THIS REPRESENTATION
not:
DOES NOT EXIST ON THE OBJECT
This may be the single most important discipline introduced by the scan/object distinction.
What an Ordinary High-Resolution Scan Usually Cannot Establish by Itself
The public digital image can suggest these questions.
It usually cannot settle them alone.
- chemical identity of inks and pigments — appearance is not composition;
- microscopic deposition sequence — which mark physically lies above another may require microscopy or specialised imaging;
- pen pressure and surface relief — flat photography captures only some indirect visual consequences;
- parchment thickness and mechanical behaviour — a two-dimensional image does not contain the full three-dimensional sheet;
- surface preparation — some scraper marks, texture and treatment traces may require closer examination;
- subsurface or erased information — ordinary visible-light images can miss material traces;
- exact translucency behaviour — show-through in an image is conditional on illumination and imaging setup;
- sewing mechanics hidden within binding structures — visible images may expose clues without exposing the complete construction;
- material identity from colour alone;
- chronological order of all interventions — sequence often requires convergent material evidence;
- tactile distinctions — texture is represented indirectly, not experienced directly;
- three-dimensional deformation — folds, cockling and page curvature can be partially visible yet incompletely measured.
These are not arguments against digital scholarship.
They are routing rules.
When the decisive variable is not encoded in the representation, more analysis of the same representation cannot create the missing variable.
More Pixels Do Not Automatically Mean More Kinds of Knowledge
Resolution is often treated as though it were a universal ladder.
Low resolution is bad.
High resolution is better.
Infinitely high resolution would therefore solve everything.
That logic is false.
More spatial resolution improves knowledge only for signals encoded spatially in that imaging modality.
If the relevant variable is chemical composition, increasing RGB pixel count does not turn visible-light photography into spectroscopy.
If the relevant variable is surface height, more flat pixels do not automatically create a full relief map.
If the relevant signal lies beneath the visible surface, magnifying the surface does not necessarily reveal it.
Resolution increases detail within a channel. It does not automatically add missing channels.
This is why “we need better scans” must be unpacked.
Better in what dimension?
Spatial resolution?
Spectral range?
Colour calibration?
Surface geometry?
Transmission?
Microscopy?
A good imaging request names the missing channel before requesting more data.
The Viewer Is Another Representation Layer
Even after an institution creates a digital master, most users do not necessarily interact directly with the master file.
They interact with a delivery system.
master capture → archival file → web derivative → zoom tile → browser → screen
The delivery layer may resize, tile, compress or otherwise transform data for efficient viewing.
Then the user may take a screenshot.
Crop it.
Paste it into a document.
Export that document as a PDF.
Upload the PDF into another system.
By the time a model or researcher analyses the image, several transformations may separate the observed pixels from the institutional capture.
This is representation distance.
DR ↑ as transformations accumulate
Distance does not invalidate a result automatically.
It changes which claims require caution.
A screenshot may be perfectly adequate for demonstrating that a large circular diagram exists.
It may be inadequate for deciding whether a one-pixel discontinuity is a physical pen lift.
Foldouts Show the Difference Between Image and Interface
The Voynich Manuscript’s foldouts make this problem unusually visible.
A digital image can present a complete foldout beautifully.
But the physical foldout is not merely a large rectangular picture.
It has an opening sequence.
Folds create layers.
Some surfaces are hidden until an action occurs.
The reader’s hands change the visible state.
Scale interacts with the body.
Yale’s 2016 facsimile project addressed exactly this representational gap. Curator Raymond Clemens explained that high-resolution images were excellent but did not fully provide a sense of the manuscript as a physical object; the facsimile was designed to convey size and structural relationships more effectively.
This gives us three different questions:
- What is drawn on the foldout? Often strongly scan-testable.
- How are the components spatially related when fully opened? Often scan/facsimile-testable.
- How does the reader physically reveal and manipulate the information surface? Better represented by facsimile or object-level interaction.
A page can be an image and an interface at the same time. A flat scan captures the first more completely than the second.
Yale News describes the 2016 facsimile and its attempt to restore book-scale and foldout relationships omitted by screen viewing: Mysterious Voynich Manuscript reborn in facsimile edition.
A Transcription Is Not a Better Scan
After the image comes another transformation.
image → transcription
A transcription is enormously useful because it converts visible marks into machine-readable symbols.
But it does not simply sharpen the image.
It changes the ontology of the data.
A human or algorithm decides that one visible form belongs to one character class. A gap becomes a separator. Two strokes become one glyph or two. An uncertain mark is encoded according to a convention. Variants may be merged.
The transcription therefore contains interpretation even before translation begins.
scan = visual measurement
transcription = visual measurement + segmentation + classification
This is why a statistical result from EVA or another transcription cannot be described as though it came directly from parchment.
The correct chain is:
object → image → transcriber decisions → symbolic corpus → statistic
The farther the result travels from the object, the more important the transformation ledger becomes.
AI Makes the Scan Problem Larger, Not Smaller
Artificial intelligence can inspect images at extraordinary scale.
It can cluster shapes.
Compare distant folios.
Recover candidate recurrence.
Measure geometry.
Generate segmentations.
But AI does not remove the imaging channel.
It consumes its output.
If the input is a cropped screenshot, the model cannot recover physical evidence that the crop never contained.
If a normalised transcription merged two visual variants, an LLM analysing the tokens cannot infer with certainty that the variants were materially identical.
If a colour-adjusted derivative changes contrast, a model may confidently classify a feature that belongs partly to the processing pipeline.
AI can amplify signal inside a representation. It cannot conjure information that the representation never carried.
This gives the AI ingestion system a mandatory field:
REPRESENTATION SOURCE / VERSION / TRANSFORMS
Without that lineage, machine-scale analysis can become machine-scale certainty about an invisible preprocessing choice.
A Practical SCAN Claim Ladder
We can now formalise how far a scan-derived claim is allowed to travel.
| Level | Authorised statement | Example |
|---|---|---|
| S0 — Image unavailable / inadequate | No claim about the proposed visual feature is authorised. | The region is missing, too blurred or too compressed. |
| S1 — Feature visible | A visual feature appears in the specified representation. | A narrow dark line is visible at this coordinate. |
| S2 — Feature reproducible | Independent or alternative image views reproduce the observation. | The same line appears in another derivative or acquisition view. |
| S3 — Geometry measurable | Size, angle, position or recurrence can be measured within known image limits. | The form occurs at line starts more often than in medial positions. |
| S4 — Representation-robust | The result survives plausible image transformations or independent transcriptions. | A recurrence remains under two segmentation regimes. |
| S5 — Object-compatible | Available object-derived evidence does not contradict the scan interpretation. | Visible colour grouping is consistent with—but does not itself prove—material analysis. |
| S6 — Object-confirmed where necessary | The decisive material question has been tested through authorised object-level evidence. | Instrumental analysis confirms that two apparently similar regions differ materially. |
The point of the ladder is not bureaucracy.
It prevents an S1 visual impression from being narrated as an S6 material conclusion.
A clear picture can still support only a small claim.
The Scan Contract
From this point forward, a serious scan-based Voynich experiment should be able to answer a compact set of questions.
- Which image source? Institutional original, facsimile photograph, derivative, screenshot or another source?
- Which folio and region? The coordinate must be reproducible.
- Which imaging modality? Visible light, spectral image, microscopy or another channel?
- What is the effective resolution? Can the proposed feature be represented at that scale?
- What transformations occurred? Crop, rotation, contrast, colour adjustment, downsampling, compression, segmentation?
- What variable is being measured? Geometry, intensity, colour, recurrence, boundary, topology?
- What variable is only inferred? Material, function, language, chemistry, chronology?
- What would the representation systematically miss?
- What is the detection floor? How small or faint can a feature be before non-detection becomes meaningless?
- What object-level evidence would change the conclusion?
A claim that cannot answer these questions may still be interesting.
It is not yet a well-specified measurement.
Padua: Which Claims Belong in the Scan Lane?
The Padua programme gives us immediate examples.
| Padua-related claim | Initial testability | Why |
|---|---|---|
| Apparent text distributions differ across sections | SCAN / transcription | The relevant evidence is visible symbol placement and encoded sequence. |
| A page-layout convention resembles a comparator | SCAN | The comparison begins in visible geometry. |
| A recurrent visual component crosses several folios | SCAN | Recurrence can be recovered visually before meaning is assigned. |
| A ruling or construction line is present | SCAN → BOTH | Visible imagery may suggest it; faint or physically ambiguous evidence may require specialised object-derived imaging. |
| One ink was applied after another | BOTH / OBJECT | Overlap appearance may generate the question; deposition sequence may require microscopy or material analysis. |
| The parchment was prepared according to a local workshop practice | OBJECT + comparators | Material preparation cannot be established from ordinary appearance alone. |
| A pigment composition localises production | OBJECT + historical controls | Chemical identity and geographic discrimination both exceed ordinary scan evidence. |
This stops a common category error.
We do not ask a distributional scan test to prove parchment origin.
We do not ask a chemical test to translate a word.
Each evidence lane gets a scientific job appropriate to the signal it can carry.
The Escalation Rule: Name the Missing Signal
Yale’s access policy effectively asks one of the best questions in Voynich methodology.
Why do you need the original?
The weak answer is:
because the original is more authentic.
The strong answer is:
because hypothesis H depends on material variable X, and X is not encoded with adequate fidelity in the available representation R.
This should become the permanent escalation rule.
- State the unresolved hypothesis.
- State the decisive variable.
- Show that existing imagery does not carry that variable adequately.
- Identify the least invasive object-derived method capable of carrying it.
- Specify what outcome would change the hypothesis.
- Preserve the object-level result as a new representation with full lineage.
This turns physical access from prestige into experimental necessity.
The object is not the reward for serious research. It is the instrument of last resort when the scientific question genuinely requires material signal.
What the Scan Gives Back to Conservation
The scan/object distinction is not a hierarchy in which the physical manuscript always wins.
Digital representation creates capabilities the unaided physical encounter does not possess.
- repeatable side-by-side comparison across distant folios;
- rapid zoom;
- precise computational coordinates;
- automated counting;
- large-scale recurrence search;
- image registration;
- statistical measurement;
- shared annotation;
- distributed replication;
- independent checking without additional handling;
- preservation of a time-stamped visual state;
- ability to test the same representation under multiple algorithms.
In some tasks, the digital representation is not merely safer.
It is scientifically better suited.
No human seated with the original can manually compare every glyph position across the entire codex as efficiently as a well-designed computational pipeline can compare a versioned dataset.
The correct principle is therefore:
Use the representation for the jobs representation does better. Use the object for the jobs the representation cannot carry.
What We Should Say Instead of “The Manuscript Shows”
Language is part of the method.
When the evidence comes from a representation, the sentence should preserve that route where it matters.
| Compressed statement | Higher-fidelity statement |
|---|---|
| The manuscript has no ruling. | No ruling is clearly visible in the specified public image set under the conditions examined; object-level or specialised imaging evidence would be needed to exclude subtle ruling more strongly. |
| These two inks are different. | The two regions differ in visible appearance; chemical identity requires instrumental evidence. |
| This glyph is one continuous stroke. | The image is consistent with one continuous visible form at this resolution; stroke production may require closer or object-derived examination. |
| The foldout is a map. | The scan preserves a large connected visual field with specified geometric relations; function remains a separate hypothesis. |
| There is no text beneath this pigment. | No underlying text is visible in this representation; absence of subsurface signal has not been established. |
The longer sentences are not weaker.
They preserve the measurement boundary.
Precision in evidence language is not hedging. It is provenance for claims.
A Scan Is a Time Slice Too
There is one more property of a scan that is easy to overlook.
It records the visible state of the object at an imaging event.
That makes a high-quality digital image a conservation witness.
Future images can be compared against earlier ones.
Changes in damage, colour appearance, deformation or surface condition can potentially be noticed—again subject to differences in imaging conditions.
S(t) = representation of object state under imaging conditions at time t
This adds another dimension to custody.
The scan is not only a substitute for physical access.
It can become part of the historical record of what the manuscript looked like under specified conditions at a particular moment.
That means future digitisation should not simply overwrite old representation states.
Versioned image generations can themselves become evidence.
The Permanent Testability Field
The series can now freeze the methodological label introduced in the custody work:
TESTABILITY(H) ∈ {SCAN, OBJECT, BOTH, CURRENTLY UNRESOLVED}
SCAN means the decisive signal is adequately encoded in a specified representation.
OBJECT means the decisive signal is materially present but not adequately represented in the available scan lane.
BOTH means representation generates or constrains the hypothesis while object-derived evidence is needed to discriminate among surviving explanations.
CURRENTLY UNRESOLVED means neither existing representations nor available object-derived evidence can presently settle the question.
The label applies to the hypothesis at its current evidence state.
It can change.
A future imaging method may move an OBJECT question into BOTH or SCAN.
A failed scan experiment may reveal that an apparently digital question depends on material variables after all.
Testability is a property of the question relative to the available evidence channels—not a permanent property of the manuscript.
World Return: A Scan Can Know a Great Deal—If We Let It Stop Where It Stops
The digital Voynich is not a lesser Voynich.
It is a different research object derived from the physical one.
It can know where visible marks are.
It can preserve geometry.
It can carry page layout across continents.
It can let thousands of researchers inspect a foldout without thousands of handling events.
It can reveal recurrence no human reader would easily notice by flipping between distant folios.
It can create stable input for computation.
It can preserve a visible time slice of the manuscript for future comparison.
That is an enormous epistemic achievement.
But the scan becomes dangerous when its success makes its boundary invisible.
It does not chemically identify ink because the pixels look black.
It does not prove a surface is smooth because relief is hard to see.
It does not prove there was no erased mark because the visible-light image contains none.
It does not turn a conventional transcription boundary into a medieval word boundary.
And it does not become the physical Voynich merely because the resolution is excellent.
A scan is trustworthy when the claim remains smaller than or equal to the signal the scan actually carried.
That may be the governing rule for the entire digital phase of Voynich research.
Use the scan aggressively.
Measure it.
Compare it.
Rotate it.
Challenge it.
But when the decisive variable falls outside the channel, stop.
Name what is missing.
Then decide whether the object deserves to be asked.
Next: Voynich | What Can Only the Object Tell Us?
Primary institutional routes: Yale Beinecke — Voynich Manuscript access, scans and original-object policy · Yale News — 2004 digitisation and 2016 facsimile.
When the Object Stayed and the Information Escaped → · The Physical Manuscript and Its Digital Double → · Custody Through Time → · Voynich Research Library →
MOVEMENT II · NEXT EVIDENCE LANE
Voynich | What Can Only the Object Tell Us? →
The inverse test: which questions depend on material variables, which object-derived measurements already exist, and what a hypothesis must prove before renewed access to Beinecke MS 408 is scientifically justified.