Voynich Corrections: Errors, Emendations and Scribal Control

One of the strangest things in the Voynich Manuscript is something that is barely there.

Correction.

A long handwritten manuscript should contain mistakes.

A copied word repeated accidentally.

A missed letter inserted above the line.

A wrong form scraped away.

A line crossed out.

A transposition mark.

An overwritten spelling.

Medieval scribes were human.

The Voynich scribes were too.

Yet the manuscript has long been famous for the scarcity and unobtrusiveness of visible corrections in its principal script.

Anne Nill, who knew Wilfrid Voynich’s papers and examined manuscript material closely, wrote in 1953 that in her experience she had never encountered a manuscript whose corrections and erasures were so unobtrusive—if it contained them at all.

She did not claim perfect certainty.

She remembered one or two probable corrections.

Later high-resolution photography changed the picture further.

Researchers can now point to a small set of places where the writing appears to have been amended.

So the strongest modern statement is not:

The Voynich Manuscript contains no corrections.

It is:

visible corrections in the principal Voynich writing are remarkably few and generally minor relative to the amount of text.

That is a more interesting result anyway.

Because now we can ask what production process makes mistakes so rare—or makes them so hard to see.

Quick Read

One-sentence answer: the Voynich Manuscript is not literally correction-free, but high-resolution inspection still finds only a small number of probable, mostly minor emendations in an unusually long principal text; this strongly suggests a fluent and controlled production process while remaining compatible with several very different mechanisms—trained copying, prepared drafting, mastered shorthand or cipher, procedural generation, or corrections accomplished by subtle stroke modification rather than conspicuous erasure.

The near-absence of corrections was noticed long before modern digital imaging.
Anne Nill’s 1953 letter is an important historical statement of the observation.
She described Voynich corrections and erasures as exceptionally unobtrusive, not necessarily nonexistent.
High-resolution images have revealed a small set of probable emendations on folios including f16r, f20v, f24v, f39r, f42r, f50v, f79r, f80r, f83r, f102v2 and f112r.
Some apparent corrections may involve adding strokes to transform one glyph into another rather than deleting the original form.
Iron-gall ink makes simple wiping difficult; scraping or overwriting are more realistic correction mechanisms.
The absence of obvious erasures therefore should not be equated with absence of error correction.
Rare visible corrections remain remarkable given the manuscript’s large amount of principal text.
Low correction rates are compatible with fluent copying from an exemplar.
They are compatible with writers who had mastered the script and its orthographic conventions.
They are also compatible with a mechanical or generative writing procedure whose legal next forms were easy to produce.
They do not prove the text is meaningless.
They do not prove the text is meaningful.
If several scribes produced the manuscript, low corrections across hands would imply that the writing conventions were teachable or shared rather than one person’s moment-to-moment improvisation.
If one hand produced most or all text, the same observation instead supports high individual fluency.
The correct model should predict the kinds of errors and corrections that the script makes easy or difficult.

The missing corrections are therefore not emptiness.

They are a production constraint.

“No Corrections” Was Always Too Absolute

The most famous version of the claim is seductive because it is clean.

No erasures. No corrections.

Historical evidence was already more careful.

Nill’s letter explicitly allows “one or two probable corrections”.

Modern high-resolution scans make subtle penwork visible that photostats could hide.

As a result, several candidate emendations are now known.

This does not destroy the original observation.

It improves it.

A manuscript containing a dozen or so plausible small corrections among tens of thousands of word-like units is still strikingly clean.

The scientific question becomes quantitative and comparative rather than absolute.

How low is the correction rate, which correction types occur, and what production processes can create that profile?

A Correction Can Be Almost Invisible

Modern readers imagine correction as deletion.

Cross out the wrong word.

Write the right one.

Handwriting offers another strategy.

Add a stroke.

One glyph becomes another.

If the Voynich alphabet contains closely related forms that differ by small additions, some mistakes can be repaired without leaving a crossed-out word.

This possibility has been noted in discussions of apparent corrections such as f20v.

It matters because the visible correction rate depends partly on the affordances of the script.

A script in which one stroke converts an illegal form into a legal neighbour can hide its repair history better than one requiring complete erasure.

Therefore:

few obvious corrections ≠ few corrected mistakes.

We need to understand correction technology before interpreting correction frequency.

Iron-Gall Ink Makes Erasure a Material Problem

The principal text was written with iron-gall ink.

This matters because correcting ink on parchment is not the same as pressing Backspace.

Iron-gall ink reacts with the writing surface.

Simply wiping the wet mark may not remove the evidence cleanly.

A scribe can scrape parchment.

Overwrite.

Add strokes.

Insert material above the line.

Or leave the mistake.

Each option leaves a different material trace.

A future correction survey should therefore distinguish:

scraped erasure;
overwriting;
stroke addition;
interlinear insertion;
deletion mark;
uncorrected anomaly.

Lumping all six into “corrections” loses the production signal.

The Known Candidate Corrections Are Small

Modern reference material collects probable emendations across a small set of folios.

They include examples on f16r, f20v, f24v, f39r, f42r, f50v, f79r, f80r, f83r, f102v2 and f112r.

The important word is probable.

Handwritten variation can resemble correction.

An odd stroke may be:

an amended glyph;
a normal allograph;
a pen hesitation;
a ligature;
a damaged stroke.

Even if every candidate is accepted, however, the broad conclusion remains.

The corrections are sparse and local.

There are no obvious pages repeatedly rewritten into final form.

No pervasive forest of editorial marks.

The principal text looks unusually fluent at the level visible to us.

Fluent Writing Does Not Tell Us What Was Understood

One historical inference from the correction scarcity was:

perhaps the scribe did not understand what was being written.

The reasoning is that a copyist can reproduce unfamiliar signs mechanically without noticing semantic errors.

This is possible.

It is not a necessary conclusion.

A trained scribe who understands the text can also write fluently.

An expert copyist can understand the script without being the author.

A cipher clerk can write encoded text correctly while understanding both plaintext and mechanism.

A generator can produce legal forms without semantics.

Visible fluency tells us about operational mastery.

It does not identify the semantic relationship between writer and text.

Copying From a Clean Exemplar Can Explain Low Corrections

Imagine the surviving Voynich Manuscript is a fair copy.

The difficult composition happened elsewhere.

A draft was revised.

Text and image positions were solved.

The final scribe copied from that prepared exemplar.

The surviving copy can then look unusually clean because it is not the first attempt.

This is historically ordinary.

Fair copies exist precisely because drafts can absorb the mess.

The drawback is evidentiary.

No original Voynich draft is known.

The exemplar hypothesis is plausible, not observed.

It remains an essential control whenever low corrections are used to argue for generation or nonsense.

A Mastered Shorthand Can Also Be Written Fluently

The Abbreviation or Alphabet article introduced another possibility.

Suppose Voynichese is a specialised shorthand or heavily abbreviated writing system.

An outsider sees strange compounds.

A trained writer sees familiar conventions.

The speed and accuracy of production can therefore be high even though modern readers find the system opaque.

Low corrections would then support competence, not secrecy or meaninglessness.

A shorthand hypothesis should make specific error predictions.

Writers may confuse similar abbreviation signs.

They may omit one minim.

They may repair a form with one added stroke.

If candidate corrections concentrate in exactly those relationships, shorthand gains support.

The error types can test the writing-system model.

A Cipher Clerk Can Be Highly Accurate Too

Ciphers are procedures.

A well-practised clerk can execute a procedure with low visible error.

If Voynich groups are generated from a table or set of substitution rules, the writing task may be more constrained than ordinary composition.

That can reduce some kinds of mistakes while creating others.

A cipher mechanism might predict:

illegal groups rarely appear;
errors tend to produce near-legal groups;
repeated plaintext creates related visible variants;
corrections modify code groups rather than prose grammar.

A real cipher theory should therefore reproduce not only correct-looking Voynich text but also its error profile.

If the proposed procedure is so cognitively cumbersome that a human should make frequent table errors, the clean manuscript becomes evidence against it.

Human executability includes error rate.

Procedural Generation Naturally Predicts Fluency

A constrained generation hypothesis has perhaps the simplest route to low corrections.

Choose a nearby legal form.

Copy it.

Change one permitted component.

Continue.

If the process defines only legal forms, the writer seldom needs to correct a linguistic mistake because there is no target sentence to violate.

Mechanical fluency therefore supports generation as one viable explanation.

But it creates another burden.

Why correct anything?

If any legal-looking variant is acceptable, an “error” should be hard to define.

Probable corrections suggest the writer sometimes had a preferred target form.

A generation model should explain what makes the uncorrected form wrong enough to amend.

The small correction set can therefore constrain generation too.

Meaningful Language Can Be Almost Error-Free

There is a tendency to treat visible mistakes as proof of meaning and clean writing as suspicious.

That is not historically sound.

A competent professional scribe can produce a clean fair copy.

A short technical phrase repeated across records can become automatic.

A writer copying familiar material may know the next word before the pen arrives.

Corrections can also happen in a lost draft.

Therefore low corrections are compatible with meaningful language.

The language model must simply explain why this particular manuscript is cleaner than many comparators.

Possible answers include fair-copy status, strong training, standardised text or specialised notation.

Those explanations can be tested against page planning and handwriting.

Multiple Scribes Would Make the Cleanliness More Interesting

The influential Lisa Fagin Davis model proposes several principal scribal hands.

Recent dissent argues that the differences may be less discrete.

Either way, the correction problem should be conditioned on hand.

If five genuinely different writers all produce similarly low correction rates, the writing system likely had shared conventions that several people could execute fluently.

That would weaken models requiring one idiosyncratic improviser whose private intuition determines every form.

If one proposed hand accounts for most candidate corrections, the difference may reveal experience level or workflow.

If correction types vary by hand while underlying word rules remain stable, the manuscript may separate personal motor habits from shared orthography.

The correction map can therefore become another independent dimension in the scribe debate.

One Hand Would Make the Cleanliness Interesting in a Different Way

Suppose future palaeography instead supports one dominant scribe.

The low correction rate then becomes a measure of individual mastery.

One person writes an unfamiliar-to-us script across a large codex with remarkable consistency.

That person either:

knew the script extremely well;
copied from a reliable source;
followed a highly constrained generation or encoding procedure;
or combined these.

The manuscript would still reject the image of a hesitant inventor making up a new alphabet word by word while repeatedly revising it on the final parchment.

The writer looks practiced.

Practice is the common denominator across both one-hand and multi-hand models.

Corrections Should Be Analysed by Error Type, Not Just Counted

A single total correction count wastes information.

Suppose most corrections add one stroke to a glyph.

That implies one kind of production error.

Suppose corrections mostly insert missing token endings.

That implies another.

Suppose entire words are scraped and replaced.

That implies something different again.

A future taxonomy might include:

glyph-shape repair;
missing-stroke repair;
token substitution;
word insertion;
word deletion;
space correction;
line-order correction;
visual-label correction.

Different mechanisms predict different error spectra.

The spectrum may be more informative than the headline scarcity.

Uncorrected Anomalies Matter Too

Correction analysis should not look only for places where the scribe changed something.

It should also identify forms that appear unusual but were left untouched.

A rare glyph may look malformed.

A repeated word may look like dittography.

An absent zodiac label may look like omission.

If the writer leaves these cases, perhaps they were acceptable.

Or perhaps the writer failed to notice them.

The distinction can help define what the production system considers an “error”.

A model that explains corrected forms but not tolerated anomalies is incomplete.

Error is a contrast class.

We need both corrected and uncorrected deviations.

Exact Repetition Creates a Special Correction Test

The manuscript sometimes writes the same token twice in succession.

In ordinary copying, accidental word duplication—dittography—is a familiar error.

If every exact Voynich double were a mistake, we might expect some to be corrected.

Many remain.

This means at least one of three things.

some doubles are intentional;
the writer tolerated certain duplications;
the writer did not recognise them as errors.

The Exact Repetition article therefore intersects correction analysis naturally.

A model that calls repetition meaningful should explain why.

A model that calls it accidental should explain why it usually remains uncorrected.

The manuscript forces both sides to pay a cost.

Word-Boundary Ambiguity May Hide Corrections

Some apparent errors may not involve glyph identity at all.

A scribe can leave too little or too much space.

Modern transcribers then disagree about whether a boundary exists.

If the original writer adjusted spacing while writing, those adjustments may not look like correction marks.

This matters especially under the 2026 preprint challenge that uncertain spaces may behave differently from strong gaps.

That work is new and not settled, but the general correction question is robust:

what does a boundary mistake look like in a script whose word boundaries are themselves under investigation?

Correction analysis must include spacing behaviour rather than only inked glyphs.

Page Planning Can Reduce the Need for Correction

The preceding production batch showed that drawings generally preceded text and writing was fitted into available spaces.

This looks difficult.

Yet the writer repeatedly fits lines around visual obstacles without leaving obvious abandoned attempts.

That suggests planning.

Perhaps a draft existed.

Perhaps the writer was adept at estimating how much material remained.

Perhaps flexible abbreviation or token generation allowed last-minute adjustment.

The page-planning and correction problems should therefore be analysed together.

A model that explains tight fit by improvisation must also explain why improvisation produces so few visible repairs.

A prepared-draft model predicts the cleanliness more naturally.

Again, the two need direct comparison.

Corrections Could Reveal the True Character Inventory

An amendment is a natural experiment.

The writer begins with Form A.

Then changes it toward Form B.

That action tells us the two forms are close enough for the writer to confuse or transform.

If repeated corrections consistently turn one rare form into one common form, the rare form may be an error rather than an independent character.

If one additional stroke repeatedly transforms one functional class into another, that stroke may carry real graphemic information.

Corrections can therefore help distinguish:

allograph;
independent grapheme;
compound;
scribal mistake.

The rare correction can be more informative about alphabet structure than a hundred perfect common forms.

Corrections Could Reveal Whether q, Benches and Minims Are Compositional

Imagine a writer accidentally produces an ok-form and adds a stroke to make qok.

That would be relevant to the q-prefix versus compound debate.

Imagine a bench is corrected by changing only the inserted gallows while the frame remains untouched.

That would support frame-plus-payload compositionality.

Imagine iin becomes iiin by one inserted minim.

That would show minim count is a dimension the writer can deliberately alter.

These examples are hypothetical.

The principle is not.

where the scribe repairs the system, the repair can expose what the scribe believed the system’s units to be.

A high-resolution correction corpus could therefore contribute directly to alphabet recovery.

Corrections Can Test a Decipherment

A decipherment should make mistakes intelligible.

Suppose a candidate corrected token is translated under a proposed language model.

The uncorrected form should produce a plausible error.

A misspelling.

Wrong inflection.

Incorrect cipher group.

Illegal abbreviation.

The corrected form should repair it.

If a decipherment gives equally fluent meanings to both pre- and post-correction forms, it has failed to explain why the scribe bothered to change anything.

This makes correction loci unusually valuable held-out cases.

They contain an implicit judgement by the historical writer:

this form was less acceptable than that form.

A real decoder should eventually recover the reason.

What Survives the Correction Work

The principal Voynich writing is remarkably clean and contains few conspicuous corrections or erasures.
The claim of literally zero corrections is too strong.
A small set of probable, mostly minor emendations is visible in modern high-resolution images.
Some correction may occur through added strokes rather than deletion.
Iron-gall ink and parchment affect how correction can be performed and detected.
Low visible correction rates imply fluent production at the operational level.
Fluency is compatible with meaningful copying, shorthand, cipher execution and structured generation.
It does not establish whether the writer understood the semantic content.
If several hands share the low correction profile, the writing rules were likely transmissible and practised.
If one hand dominates, that individual was highly practised.
Correction types can help recover grapheme boundaries and production rules.
A successful decipherment should explain why corrected forms are better than the forms they replace.

What Does Not Survive as Established Knowledge

The Voynich Manuscript has absolutely no corrections.
Low correction rate proves meaningless text.
Low correction rate proves meaningful text.
The scribes did not understand what they wrote.
The scribes fully understood every semantic layer.
Every odd glyph is a correction.
Every doubled token is a copying error.
Every error could be repaired invisibly.
One proposed hand definitely made more mistakes than another without a complete controlled survey.
A clean final manuscript means there was no draft or exemplar.

The exceptional fluency survives.

The explanation remains open.

A Better Correction Analysis

Build a manuscript-wide high-resolution correction catalogue.
Separate principal text from later marginalia and numbering.
Classify scraping, overwriting, stroke addition, insertion and deletion separately.
Record uncertain cases rather than forcing correction status.
Map each case to proposed hand, Currier regime, bifolium and document role.
Compare corrected forms with their closest word-family neighbours.
Catalogue uncorrected anomalies as a control.
Test whether error types match proposed glyph composition.
Compare correction rates with period manuscripts under matched page and token lengths.
Require language, cipher and generation models to predict the observed error spectrum.
Use correction loci as validation cases for decipherment.

What Would Count as a Real Correction Breakthrough?

Imagine every probable correction is independently reviewed from high-resolution images by palaeographers.

A stable corpus of true emendations emerges.

Most errors fall into a few recurring transformations.

For example:

one missing stroke;
one wrong minim count;
one confused bench component;
one illegal token ending.

The transformations align with a compact character grammar derived independently from ordinary text.

Then a decipherment predicts why each pre-correction form is wrong and why the repaired form is valid.

That would be a major breakthrough.

The correction marks would stop being curiosities.

They would become the scribes’ own evidence about the rules they were following.

A correction is one of the rare places where the historical writer tells us, without translation, that one visible form was preferable to another.

Primary School: Fix the Letter Without Erasing It

Write an invented symbol.

Then add one stroke so it becomes another legal symbol.

Ask the child whether the first mistake has disappeared.

Not completely.

It has been absorbed into the corrected form.

This demonstrates why a manuscript can contain correction without crossed-out words.

Lower Secondary: What Kind of Error Is It?

Give students examples of:

misspelled word;
duplicated word;
wrong spacing;
wrong letter;
missing word.

Ask how each would be corrected on parchment.

Different errors leave different physical evidence.

The exercise teaches why “correction count” is less informative than correction type.

Upper Secondary: Compare Two Production Models

Model A copies meaningful prose from a clean exemplar.

Model B generates legal pseudo-words by a simple procedure.

Both can produce few corrections.

Ask what kinds of mistakes each should make.

Copying may produce omissions, repetitions and substitutions against a target.

Generation may produce illegal forms or procedural slips but has no semantic target to violate.

The error spectrum, not just the count, can distinguish them.

JC and Adult Readers: Corrections as Counterfactual Data

At a higher level, a correction gives us two nearby data points.

The form that was initially produced.

The form the scribe preferred afterward.

This is a natural counterfactual pair.

A model can ask:

Which has higher probability under the writing grammar?
Which fits the latent syntax better?
Which decrypts legally?
Which belongs to the expected word family?

If the historical correction consistently moves from low-probability to high-probability form under one model, that model gains powerful independent support.

Correction loci may therefore be disproportionately valuable despite their rarity.

A Parent and Teacher Guide

Replace “no corrections” with “very few visible corrections”.
Remember that adding a stroke can repair a glyph.
Separate ink technology from cognitive explanation.
Do not equate fluent writing with semantic understanding.
Compare fair-copy, shorthand, cipher and generation models.
Study which mistakes were corrected and which were tolerated.
Use the correction to test the alphabet and decipherment.

The general reasoning lesson is unusually strong:

absence of visible repair is evidence about a production process, but the process must be identified before the absence becomes a semantic conclusion.

Reader Checklist: Before You Use “No Corrections” as a Voynich Argument

Are you claiming zero corrections or an unusually low correction rate?
Have high-resolution candidate emendations been considered?
Could the glyph have been repaired by adding strokes?
Could scraping or overwriting be subtle?
Does the apparent correction belong to the principal hand layer?
What type of error was corrected?
What comparable anomaly was left uncorrected?
Does the rate change by proposed scribe?
Does it change by Currier regime?
Could a clean exemplar explain it?
Could procedural generation explain it?
What errors should the proposed cipher make?
Does the decipherment explain why the corrected form is better?

Frequently Asked Questions

Does the Voynich Manuscript really have no corrections?

No. Modern high-resolution inspection has identified a small set of probable emendations. The stronger observation is that visible corrections are unusually sparse and generally minor.

Why was it described as correctionless?

Earlier researchers working from the manuscript and photographic copies were struck by how difficult corrections and erasures were to find. Anne Nill documented this impression explicitly in 1953 while still allowing one or two probable cases.

Where are probable corrections found?

Modern reference material illustrates candidate emendations on a small set of folios including f16r, f20v, f24v, f39r, f42r, f50v, f79r, f80r, f83r, f102v2 and f112r.

Could corrections be hidden by the script?

Yes. Some related Voynich glyphs can potentially be transformed by adding strokes, so an amendment may leave less obvious evidence than deletion and rewriting.

Does low correction rate prove meaningless generation?

No. It is compatible with generation but also with fluent copying, a clean exemplar, mastered shorthand, technical notation or disciplined cipher execution.

Does it prove the scribe understood the text?

No. It demonstrates operational fluency more directly than semantic understanding.

Why are corrections useful for decipherment?

They give two related forms and show which one the historical writer preferred. A good writing-system or decipherment model should explain why the corrected form is more appropriate.

What is the strongest current conclusion?

The principal Voynich script was produced with remarkable visible fluency. The manuscript is not literally correction-free, but its sparse emendation profile is a major constraint on any model of how the text was copied, encoded, generated or understood.

Research and Further Reading

The Final Idea

The almost-correctionless Voynich Manuscript does not look hesitant.

Its writers move through the script as though the system is familiar.

Lines fill awkward spaces around drawings.

Thousands of token forms obey strong internal constraints.

Rarely, a stroke seems to be changed.

A form is repaired.

Those tiny repairs may be more valuable than their rarity suggests.

They show that the writer had preferences.

Some visible states were better than others.

There were rules worth restoring.

We do not yet know whether those rules were linguistic, cryptographic, scribal or generative.

But we know the pen was not wandering randomly across the parchment.

The most important fact may not be that Voynich contains few corrections. It may be that the few corrections it does contain are places where the manuscript briefly reveals what its writer considered wrong.

Voynich | Everything eduKate Knows and Tested | The Almost-Correctionless Manuscript: Errors, Emendations and Scribal Control