Voynich and Padua | The Preregistration Freeze: Lock the Localisation Rules Before Looking at the Voynich Result

There is a moment when research has to stop learning from its target.

For the Voynich and Padua programme, that moment comes after calibration, blind localisation, false-positive testing and out-of-sample validation.

At that point the localisation framework has earned a version.

And the version must be frozen.

The Preregistration Freeze is the point at which the localisation rules are timestamped, versioned and locked before the validated system is allowed to see its final Voynich result.

This is not bureaucracy.

It is how we stop the unknown manuscript from teaching the method how to identify the unknown manuscript.

Why Freeze After Validation?

The localisation system now contains many adjustable parts:

candidate regions;
Coordinate Stack definitions;
feature extraction rules;
evidence-family weights;
dependence discounts;
expected-evidence penalties;
false-positive thresholds;
minimum-path costs;
null-coordinate threshold;
tie-breaking rules;
confidence language.

If those parts can change after seeing the Voynich score, any result can be rehabilitated.

A weak Padua result can become stronger by increasing codicology weight.

A strong Vienna result can be weakened by calling a parameter imported.

A null result can be converted into a preferred candidate by lowering the assignment threshold.

All of those changes might be defensible in a future model.

They cannot be allowed to rewrite the frozen run that already happened.

What Preregistration Does

Preregistration is an established open-science practice in which hypotheses, methods and analysis plans are specified before results are known. The Center for Open Science describes registrations as timestamped, read-only snapshots that can preserve a research plan before analysis.

The Voynich application is slightly unusual because the object already exists and has been studied for centuries.

What must be hidden is not the manuscript itself.

What must be frozen is the decision system before the final validated localisation run is interpreted.

The Freeze Has to Include More Than the Final Formula

Freezing only a final score equation is not enough.

The pipeline begins much earlier.

Layer to freeze	Why it matters
Candidate-region set	Prevents adding a new region only after the original candidates perform poorly.
Target Coordinate Stack level	Prevents switching from physical production to intellectual ancestry after seeing the result.
Feature definitions	Prevents redefining ambiguous features to favour a preferred candidate.
Evidence weights	Prevents post-hoc weight tuning.
Dependence rules	Prevents correlated clues from being split or merged opportunistically.
Expected-evidence penalties	Prevents missing predictions from being ignored after failure.
Minimum-path penalties	Prevents adding free handoffs to rescue the preferred geography.
Null threshold	Prevents turning an unresolved result into a forced assignment.
Reporting language	Prevents “compatible” becoming “identified” after the fact.

Freeze the Voynich Packet Too

The final input packet must also be fixed.

If one analyst can keep adding new folios, new measurements or new interpretive variables until a desired region wins, the method is not frozen in practice.

The packet should declare:

which folios and production units are included;
which scribal-hand assignments are used;
which codicological observations are admissible;
which visual features are neutral enough to enter;
which semantic claims are excluded;
which uncertain observations are represented as ranges rather than point values.

The Freeze Separates Confirmatory From Exploratory Work

Freezing does not forbid new ideas.

It labels them correctly.

After the frozen run, a surprising result may reveal a new feature worth investigating.

Excellent.

That becomes exploratory work for version 2.

It does not modify version 1.

Exploration is allowed to learn from the Voynich result. Confirmation is not allowed to pretend it did not.

A Freeze Needs a Timestamp and an Immutable Record

A useful freeze should generate a durable receipt.

version ID;
date and time;
feature-manifest version;
weight table;
candidate regions;
validation performance;
false-positive matrix;
null threshold;
code or rule hash where applicable;
signed or read-only archive location.

The Open Science Framework supports timestamped registrations and version control; reproducibility frameworks such as ENCORE likewise emphasise structured documentation and versioned computational records.

The exact platform matters less than the principle:

someone reading the final Voynich result should be able to reconstruct exactly which method existed before the result was seen.

The No-Retroactive-Threshold Rule

Suppose the frozen Padua assignment threshold is 0.80 and the Voynich returns 0.74.

The result is below threshold.

We are not allowed to decide afterward that 0.70 would have been more sensible.

We may design a new version with a 0.70 threshold if fresh validation supports it.

But version 1 remains a non-assignment.

The No-Retroactive-Feature Rule

Suppose the final run reveals that one unscored image feature strongly resembles a Paduan comparator.

Interesting.

It cannot be inserted into the completed run.

It becomes a candidate feature for the next development cycle, which must then be recalibrated on known manuscripts before Voynich is run again.

The No-Retroactive-Candidate Rule

Suppose none of the frozen candidate regions fits well, and a new region suddenly looks attractive after the result.

The correct version-1 output may be “other / unresolved.”

The new candidate enters version 2 only after known-manuscript calibration.

This protects the Null Coordinate from being treated as a temporary inconvenience.

The Freeze Is Especially Important Because We Already Know Voynich Too Well

True blindness to Beinecke MS 408 is impossible for researchers who have spent years studying it.

We know the plants.

We know the nymphs.

We know the rosettes.

We know the five-scribe discussion.

We know the competing provenance stories.

The best substitute for impossible innocence is procedural commitment.

We cannot unknow the manuscript.

We can freeze what we are allowed to change after the result.

The Preregistration Freeze Connects the Whole Validation Ladder

Calibration tells us whether scores match known outcomes.
Blind localisation reduces answer contamination.
False-positive testing tells us how often Padua is overcalled.
Out-of-sample testing tells us whether the rules generalise.
Preregistration Freeze prevents the validated system from changing after it encounters its final target.

The eduKate Freeze Protocol

Choose the validated method version.
Freeze candidate regions and coordinate definitions.
Freeze the complete evidence-feature manifest.
Freeze weights, dependence discounts and missing-evidence penalties.
Freeze minimum-path costs and null thresholds.
Freeze the Voynich input packet.
Generate a timestamped immutable receipt.
Run the localisation once.
Archive raw scores and intermediate outputs before interpretation.
Report the frozen result even if inconvenient.
Move every post-result idea into a new exploratory version.

What Counts as a Breach?

changing weights after seeing the Voynich region ranking;
adding or removing an evidence family because it changes the winner;
lowering an assignment threshold after a near miss;
adding a new candidate region to avoid a null result;
reclassifying a failed prediction as irrelevant after the run;
reporting only the tuned rerun and hiding the frozen first result.

Any of those actions may be legitimate exploratory research.

They cease to be legitimate when presented as the original frozen test.

Research Sources

The Final Idea

The final Voynich test should be allowed to surprise us.

That can happen only if the method is no longer allowed to move out of the surprise’s way.

Freeze the compass before entering the fog.