Voynich Noise Problem: Which Differences Carry Information?

Voynich research is full of differences.

One glyph is taller.

One token has an extra stroke.

One space is narrow.

One page prefers chedy-like forms.

One scribe appears to draw a gallows differently.

One line ends with a form rarely found elsewhere.

Every one of those differences can tempt interpretation.

New letter.

New word.

New grammar.

New cipher state.

New meaning.

But some differences may carry no semantic information at all.

They may be handwriting.

Allography.

Spelling tolerance.

Operator choice among equivalent cipher forms.

Line fitting.

Damage.

Modern transcription disagreement.

The Noise Problem asks which observable differences change the underlying information state—and which differences can vary while the intended message remains the same.

Quick Read

Direct answer: Voynich contains both strong structured variation and genuine uncertainty about what counts as a meaningful distinction. Noise may enter through handwriting, allographs, abbreviation, cipher homophones, operator preference, spelling variation, damage, ambiguous spaces and modern transliteration. The correct goal is not to remove variation aggressively. It is to identify transformations under which the manuscript’s deeper structure remains stable, then treat the residual differences as candidate information-bearing signal.

Noise does not mean meaningless manuscript.
Noise means variation that does not change the underlying information state relevant to the model being tested.
A glyph difference can be meaningful at one level and noise at another.
Two handwriting variants can encode the same functional unit.
Two homophonic ciphertext forms can encode the same plaintext unit while remaining deliberately different on the surface.
A narrow versus wide space can be layout variation or a real boundary distinction.
Modern transcriptions introduce their own uncertainty.
The Almost-Correctionless article owns visible emendations; this article owns unexplained variation whether corrected or not.
The Drift article owns systematic change across manuscript order; this article owns local signal-versus-noise classification.
Currier A/B is too coherent to be dismissed as noise wholesale.
Some within-Currier variation may still be allographic or operational noise.
Recent 2026 work argues that some uncertain spaces form a measurable weaker boundary class rather than random transcription error.
Recent workshop-cipher modelling shows that modest operator variability can help produce realistic surface distributions without changing the underlying mechanism.
A good noise model should improve prediction after normalisation without destroying known structural effects.
No single canonical Voynich “denoising” scheme is currently accepted.

Noise Is Relative to the Question

Suppose two scribes write the same letter A differently.

For lexical identity, the shape difference may be noise.

For palaeographic attribution, the same difference is signal.

This is crucial.

There is no universal bucket labelled “noise”.

There is only variation irrelevant to one target variable.

Therefore every denoising operation should state:

what distinction is being collapsed;
for which research question;
what evidence says the distinction is functionally equivalent.

Handwriting Is the Obvious Noise Source

No human draws the same shape identically every time.

Stroke width changes.

Slant changes.

Loops open and close.

Pen angle changes.

A rushed form differs from a careful one.

If every visual difference becomes a separate character, the alphabet explodes.

If too many differences are merged, real functional distinctions disappear.

The Alphabet Problem is therefore partly a noise-classification problem.

Allographs Are Structured Noise at One Level

An allograph is a different written form of the same underlying character.

English print has uppercase A and lowercase a.

The shapes differ radically.

The alphabetic identity is related.

Voynich has candidate allographic relationships among gallows, benches, rare variants and minimally different forms.

But allography must be demonstrated by distributional interchangeability, palaeographic transformation or historical comparison.

Visual resemblance alone is not enough.

Cipher Homophones Are Deliberate Surface Noise

In a homophonic cipher, several ciphertext forms can represent one plaintext unit.

The visible variation is intentional.

Its purpose may be to prevent frequency analysis.

At the plaintext level, the difference between the homophones is noise.

At the cryptographic level, their choice distribution is signal about the key and operator.

This is why the Workshop-Cipher and Homophony articles matter.

Denoising too early could erase the very homophone structure needed to recognise the mechanism.

Operator Preference Is Neither Meaning nor Randomness

A writer can prefer one legal form over another.

That choice may not change plaintext.

Yet it can create stable statistical fingerprints.

Recent 2026 workshop-cipher modelling explicitly uses modest spelling or operator variability as part of a generative channel.

This matters because visible variation can be reproducible without being semantic.

Regularity does not prove meaning.

It can prove habit.

Spelling Variation Is Common in Historical Texts

Modern standard spelling trains us to treat one word as one fixed string.

Medieval and early modern writing is less rigid.

A name can have variants.

An abbreviation can change by scribe.

A vernacular can tolerate several spellings.

Therefore near-neighbour Voynich forms need not be separate lexical entries.

But they also need not be spelling variants.

Word-family structure keeps both possibilities alive.

Line Fitting Can Create Nonsemantic Variation

Voynich lines behave differently at their beginnings and endings.

Some words are longer at one edge.

Some forms prefer specific positions.

Part of that structure may be linguistic or cryptographic.

Part may help fit text into available visual space.

If an optional terminal element is used to fill a line, its variation is not necessarily lexical.

But line fitting is itself a production rule and can be measured.

Spaces Are a Noise Problem Only Until They Are Measured

Voynich transcribers have long marked some gaps as uncertain.

A naive view says those are human annotation noise.

Recent 2026 work reports that uncertain spaces behave systematically differently from ordinary spaces and are physically narrower on the page.

If replicated, that is important.

The uncertainty may be revealing a real graded boundary system.

What looked like transcription noise becomes manuscript signal.

Transcription Creates a Modern Noise Layer

The Voynich text we analyse computationally is not the parchment itself.

It is a model of the parchment.

Transcribers decide:

where one glyph ends;
whether two strokes form one compound;
whether a mark is damaged;
whether a space is real;
how a rare form is represented.

Different transcription conventions can therefore inject differences that never existed as functional distinctions for the historical reader.

Any fragile statistic should be repeated under several reasonable transcriptions.

Damage Can Masquerade as Rare Information

A faded stroke can create a new apparent glyph.

A stain can hide one component.

Contact transfer can introduce a misleading mark.

Physical evidence should therefore be consulted before rare forms receive semantic weight.

The Damage and Repair and Contact-Transfer articles own the physical mechanisms.

This article owns their downstream effect on signal classification.

Visible Corrections Are Only One Kind of Error

The Almost-Correctionless Manuscript article asks why explicit emendations are so rare.

But an uncorrected deviation can still exist.

A trained writer may make a variant that remains understandable.

A cipher operator may choose a legal alternative.

A shorthand writer may omit a predictable element.

The scarcity of corrections therefore does not imply zero noise.

Currier A/B Is Too Large to Call Noise

One temptation is to “normalise away” Currier differences as scribal spelling.

That would be premature.

A/B differences are broad, reproducible and visible across several resolutions.

They are signal at the level of text-state classification.

Whether that signal represents language, dialect, cipher, shorthand, notation or production remains unresolved.

Denoising should never erase a reproducible state simply because its semantics are unknown.

Drift and Noise Are Different Axes

Noise asks whether two variants encode the same underlying state.

Drift asks whether the underlying state itself changes through production.

One system can have both.

A spelling convention drifts gradually while each local state still contains harmless allographic variation.

Separating them prevents one from being mistaken for the other.

Noise Can Be Deliberate

Cryptography creates the most counterintuitive case.

Randomness can be part of the intended design.

Choose among homophones.

Insert a null.

Vary spelling.

Avoid immediate repetition.

The surface becomes noisier to protect the source.

Therefore “noise” cannot be equated with accidental error.

Noise Can Also Carry Meta-Information

Two equivalent ciphertext forms may reveal which scribe wrote them.

One abbreviation preference may reveal region.

One allograph may reveal production phase.

So a variant can be semantically noisy but historically informative.

Good analysis keeps multiple information layers separate instead of deleting the variant permanently.

The Safest Denoising Is Reversible

Never destroy the original transcription.

Create a normalised analytical layer.

Merge candidate allographs.

Collapse homophone candidates.

Reclassify weak spaces.

Then compare results with the unnormalised text.

If one conclusion survives several plausible normalisations, it is stronger.

A Useful Noise Model Should Improve Prediction

Suppose merging two glyph variants is correct.

What should happen?

word families may become cleaner;
rare-type inflation may decrease;
held-out token prediction may improve;
Currier structure should remain if it is deeper than the allograph;
source phonotactics may become more coherent.

If normalisation only makes one preferred theory look prettier while damaging other independent signals, the merge is suspect.

Noise Should Be Estimated, Not Assumed

One approach is to build explicit error or variation channels.

How often can one glyph become another?

How often is a weak space inserted?

How often does an operator choose a homophone?

Then fit those rates on one corpus and test them elsewhere.

Noise becomes a model parameter with a failure condition.

What Survives the Noise Work

Not every visible Voynich difference needs to carry lexical or semantic information.
Handwriting, allography, spelling variation, cipher homophony and operator choice can create structured surface variation.
Damage and transcription add later noise layers.
Some uncertain spaces may reflect a real graded boundary distinction rather than mere transcription uncertainty.
Currier A/B remains strong signal and should not be normalised away casually.
Noise and drift are distinct variables.
Deliberate cryptographic noise can preserve source meaning while changing the surface.
A variant can be noise for semantics but signal for palaeography or production history.
Denoising should be reversible and tested against held-out prediction.
No accepted universal Voynich noise model exists.

What Does Not Survive as Established Knowledge

Every rare glyph is a new meaningful character.
Every spelling variant is a different word.
Every uncertain space is a transcription mistake.
All Currier differences are scribal noise.
All within-page variation is meaningful.
The absence of visible corrections means the manuscript contains no errors or noise.
Denoising that improves one statistic proves the normalisation is correct.
Deliberate noise proves cipher.

A Better Noise Analysis

State which target information layer is being protected.
Separate palaeographic, lexical, cryptographic and transcription variation.
Preserve the raw representation.
Create reversible normalisation hypotheses.
Estimate variant transition rates on one bounded corpus.
Test whether normalisation improves several independent structural measures.
Repeat under multiple transcriptions.
Check that Currier, line and document-role signals are not erased unintentionally.
Validate on unseen bifolia.
Keep variants that remain functionally distinguishable.

What Would Count as a Real Noise Breakthrough?

Imagine two visually different glyph families are suspected allographs.

Palaeography shows a plausible continuous transformation.

The two forms occupy indistinguishable token slots after controlling hand and Currier state.

They never contrast in matched lexical environments.

Merge them under a frozen rule.

Held-out token prediction improves.

Word-family inflation falls.

Independent Currier and edge structures remain intact.

Then an external historical script supplies the same allographic relation.

That would be a genuine noise classification breakthrough.

We know a Voynich difference is noise when collapsing it removes complexity without removing predictive information.

Primary School: Same Letter, Different Handwriting

Ask five children to write the letter A.

The shapes differ.

The letter identity can remain the same.

This shows why visible difference does not automatically mean different information.

Secondary School: Add Deliberate Noise

Encrypt a short message using two legal symbols for each common letter.

Choose randomly between them.

The ciphertext varies while the plaintext remains fixed.

Students see how deliberate surface variability can coexist with stable meaning.

JC and Adult Readers: Noise as a Channel Model

At a higher level, observed Voynich forms can be modelled as latent functional units passed through several variation channels.

Handwriting channel.

Orthographic channel.

Cryptographic choice channel.

Transcription channel.

The challenge is identifiability: collapsing too little leaves noise inside the signal; collapsing too much destroys genuine distinctions.

Reader Checklist: Before You Call a Voynich Difference Meaningful

Could the difference be handwriting?
Could it be an allograph?
Could it be damage?
Could it be transcription convention?
Could it be a legal cipher homophone?
Could line position explain it?
Does it contrast in matched contexts?
Does the distinction survive hand and Currier controls?
Does preserving the distinction improve held-out prediction?
What information is lost if the forms are merged?

Research and Further Reading

The Final Idea

Voynich may contain more information than we think.

It may also contain more harmless variation than we think.

The danger lies in choosing too early.

Treat every difference as meaning and the dictionary explodes.

Treat every difference as noise and the real grammar disappears.

The correct model should compress only the distinctions the manuscript itself allows us to collapse.

The Noise Problem ends when variation can be normalised without losing the structures that predict what comes next—and the remaining differences become harder, not easier, to explain away.

Voynich | Everything eduKate Knows and Tested | The Noise Problem: Which Differences Carry Information and Which Are Just Variation?