Voynich and Padua | The Evidence Independence Problem: When Five Clues Are Really One Clue Repeated Five Times

A hypothesis can look stronger merely because we describe the same evidence several times.

The plant looks Italian.

The drawing style looks northern Italian.

The herbal section resembles Italian herbals.

The botanical tradition feels Paduan.

The page resembles the Carrara medical-image world.

That can sound like five clues.

It may be one visual judgement repeated five ways.

The Evidence Independence Problem is the risk of counting correlated observations as if they were separate confirmations.

For a research programme as large as the Voynich estate, this is not a minor statistical technicality.

It is now one of the main safeguards against accidental certainty.

Why a Large Research Estate Is Especially Vulnerable

We now have dozens of Padua articles, manuscript controls, network biographies, visual comparators, codicological tests and localisation arguments.

That scale is useful.

It also creates the illusion that many articles automatically mean many independent pieces of evidence.

They do not.

Ten articles may all descend from one fact.

A medical network.

A radiocarbon window.

A resemblance to one image family.

A single catalogue attribution.

Article count is not evidence count.

Three Kinds of “Many Clues”

Case	What it looks like	What it really is
True convergence	Codicology, documentary provenance and a local astronomical parameter all independently point to the same place.	Several independent evidence families.
Partial dependence	Several image features all descend from the same copied exemplar.	Multiple observations, one ancestry channel.
Re-description	“Italian-looking plant,” “Italian herbal style,” “northern-Italian botanical image.”	One judgement stated several ways.

The first can legitimately accumulate weight.

The second needs discounting.

The third should usually count once.

The Shared-Cause Test

Before counting two clues separately, ask:

Could one underlying historical cause produce both observations?

If yes, the clues are at least partly dependent.

Example:

a plant image resembles a Paduan exemplar;
the caption placement also resembles the same exemplar;
the page order resembles the same exemplar.

Those may all be explained by one copying event.

That is still important evidence.

But it should not be treated as three independent roads to Padua.

The Same-Source Test

Two claims are not independent if they ultimately rely on the same source record.

Suppose one catalogue entry says a manuscript was made in Padua in 1431.

We then write separate articles about its surgery, anatomy, pharmacy and poison text.

Those are several observations about the manuscript’s capabilities.

But the geographic coordinate “Padua 1431” still comes from one catalogue attribution.

We must not silently count Padua four times.

The Same-Measurement Test

Different analyses can also reuse the same measurement.

Radiocarbon dating may support the early fifteenth-century material window.

Several historical articles may then use that window to investigate Padua, Vienna or Prague.

The historical investigations are different.

The radiocarbon evidence remains one evidence family.

Repeated reuse increases relevance, not independent support.

The Same-Observer Test

Subjective visual classifications can be correlated because the same researcher made them.

If one observer sees “Italian” in plants, diagrams and layout, that is not automatically three independent judgements.

Independent coding, blinded comparison and pre-defined features become especially valuable here.

This is one reason the Voynich programme freezes observations before semantic naming whenever possible.

Independence Does Not Mean Unrelated

Perfect statistical independence is often unrealistic in historical research.

The goal is not to pretend evidence families never interact.

The goal is to know how they depend on one another.

Codicology and binding can be related.

Script and scribal training are related.

Image ancestry and page layout can be related.

Documentary ownership and later paratext can be related.

We should model those dependencies rather than count every visible output as a fresh vote.

A Better Evidence Ledger

Claim	Evidence family	Underlying source/cause	Independent?
Parchment dates to early 15th c.	Material dating	Radiocarbon measurements.	Independent of visual resemblance.
Several images resemble a known herbal lineage.	Visual genealogy	Potential shared exemplar.	Partly dependent internally.
One hand resembles a regional scribal tradition.	Palaeography	Scribal training.	Potentially independent of image ancestry.
A fifteenth-century inventory names the actual object.	Documentary provenance	Contemporary record.	Highly independent of visual and material evidence.

The ledger should record not only what the evidence says, but whether it adds new information.

Why Independence Matters More After the Localising Residue

The Localising Residue deliberately removed widely shared medieval features from the Padua localisation score.

The evidence that remains is rarer.

That makes double counting more dangerous.

If three surviving “rare” clues all derive from one source family, our residue is smaller than it looks.

Therefore the residue needs an independence audit before it can accumulate geographic weight.

The Coordinate Stack Helps Prevent Double Counting

The Coordinate Stack separates knowledge ancestry, text origin, image ancestry, scribe training, physical production, assembly, binding, ownership and later reception.

This helps because two clues assigned to the same coordinate and same causal mechanism deserve more scrutiny than two clues from genuinely different coordinates.

For example:

three plant-image similarities may all belong to C3 image ancestry;
a scribal-training signal belongs to C4;
a quire-production habit belongs to C5;
a dated inventory belongs to C8.

The second group is structurally closer to independent convergence.

The eduKate Independence Audit

List every clue supporting the hypothesis.
Assign each clue to an evidence family.
Record the primary source or measurement underneath it.
Ask whether two clues could be caused by one copying, production or transmission event.
Merge simple re-descriptions into one clue.
Discount partially dependent observations.
Separate researcher judgement from externally measured evidence.
Prefer convergence across material, palaeographical, documentary and local-parameter families.
Do not allow article count to substitute for evidence count.
Recalculate the localisation picture after dependency reduction.

What Would Strong Independent Convergence Look Like?

Imagine a future state in which:

a codicological study independently places a production unit in a narrow north-Italian tradition;
a blinded palaeographical study independently identifies a compatible scribal-training environment;
a recovered astronomical parameter independently matches Padua better than rival centres;
a documentary record independently places the actual manuscript in a Paduan collection soon after production.

Those would not be four descriptions of one resemblance.

They would be four different evidence channels converging on related coordinates.

That is the standard worth aiming for.

What This Does Not Mean

Correlated evidence is not worthless.
Several features within one image family can still establish that family strongly.
A large research estate is not a problem if dependencies are explicit.
The Padua hypothesis does not need every clue to be independent.
Independence is not a reason to discard coherent historical networks.

The rule is narrower:

Evidence may reinforce one mechanism without multiplying the number of independent mechanisms.

Research Links

The Final Idea

A theory should not become five times stronger because we found five ways to describe one attractive observation.

The decisive Voynich breakthrough will be strongest when different evidence systems arrive at the same conclusion for different reasons.