The easiest way to find Padua is to know in advance that Padua is the answer.
That sounds obvious.
It is also one of the hardest biases to remove from manuscript research.
A shelfmark leads to a catalogue.
The catalogue gives a city.
The city activates a network of familiar scholars, scripts, workshops and institutions.
Then every ambiguous feature begins to look compatible with what we already know.
The Blind Localisation Test reverses the order: hide the answer first, freeze the evidence packet, make the localisation judgement, and reveal the catalogue coordinate only afterward.
This is not because catalogue data are untrustworthy.
It is because the catalogue answer should be the test target, not a clue available during the test.
Blind Analysis Exists to Break the Feedback Loop
Blind analysis is used in scientific fields where researchers worry that knowing the desired or expected result may influence choices made during analysis. The basic idea is to conceal outcome-relevant information until methods, selections or cuts are fixed.
For manuscript localisation, the equivalent outcome information includes:
- shelfmark;
- library name;
- catalogue production city;
- explicit colophon;
- known owner;
- famous named author whose biography reveals the region;
- modern article title containing the location.
Those fields can be restored after the inference packet is sealed.
What Should Remain Visible?
A blind packet cannot remove everything.
The method still needs evidence.
The challenge is to preserve features that could legitimately localise the object while removing the answer label itself.
| Keep visible | Hide initially |
|---|---|
| Material substrate and dimensions | Library and shelfmark |
| Ruling, pricking, quire structure | Catalogue city |
| Script images and neutral palaeographic features | Scribe name if it reveals geography |
| Ink/pigment observations | Known owner and provenance chain |
| Image topology and page architecture | Modern descriptive captions that name the region |
| Textual structure in neutral form | Author names when the goal is physical-production localisation |
| Dating evidence | Answer-bearing colophons until reveal stage |
There Are Several Levels of Blinding
One blind test is not enough for every question.
- Level B1 · Label blind: hide shelfmark, library and catalogue place.
- Level B2 · Network blind: also replace famous author names with neutral identifiers.
- Level B3 · Text-semantic blind: describe contents only by functional type—diagnostic table, recipe block, body image—without the historical title.
- Level B4 · Image-source blind: compare visual structures without showing which comparator family they belong to.
- Level B5 · Full analyst blind: a separate analyst receives only the frozen feature packet and candidate-region definitions.
The stronger the blinding, the more we learn about whether the result is driven by the evidence rather than recognition.
Blind Physical Production Separately From Intellectual Ancestry
The Coordinate Stack matters again.
If the target is C5 physical production, named authors are often contaminating clues because their biographies reveal intellectual geography rather than where this copy was made.
If the target is C2 text origin, those same author names may be legitimate evidence.
Blinding must therefore be coordinate-specific.
Otherwise the test may “succeed” by answering a different question.
A Blind Test Needs a Candidate Set Before the Packet Is Opened
Do not let the analyst invent candidates after seeing the manuscript.
For a first early-fifteenth-century medical calibration, the candidate set might be:
- Padua / north-eastern Italy;
- Vienna / Austrian region;
- southern Germany;
- Bohemia / Prague region;
- Bologna / Emilia;
- Pavia / Lombardy;
- other / unresolved.
The exact resolution should be fixed in advance.
Otherwise a wrong city can be converted after the fact into a “close regional success.”
The Analyst Must Also Be Allowed to Say Null
A forced-choice blind test can create false confidence.
If the packet does not discriminate, the correct answer may be:
insufficient evidence to localise at this resolution.
The Null Coordinate therefore belongs inside the blind test, not outside it.
The Blind Packet Must Be Frozen
If the analyst can request more clues after forming an impression, the blind procedure begins to leak.
Therefore each packet should have a manifest:
- included images;
- included measurements;
- included text descriptors;
- excluded answer-bearing fields;
- candidate set;
- scoring rules;
- confidence scale;
- null threshold.
The packet is then hashed or versioned before analysis.
Any later addition becomes a new test version, not a silent patch.
Why Pal. lat. 1299 Is an Excellent Blind Control
Pal. lat. 1299 is dangerous precisely because its contents are so compatible with the Paduan medical world.
Anatomy.
Surgery.
Pharmacy.
Multilingual terminology.
If our blind method sees those features and calls Padua, it has learned the wrong features.
The known catalogue answer—Vienna 1413–1414—becomes a useful trap for our own assumptions.
Why Pal. lat. 1265 Is an Even Harder Test
Pal. lat. 1265 should not produce one clean city.
A method that confidently forces Padua or Germany is failing in a different way.
The desired output is a mixed or unresolved production state at the resolution permitted by the evidence.
Blind testing therefore evaluates not merely whether we guess correctly, but whether we preserve complexity when the answer is complex.
The Reveal Must Be One-Way
Once the catalogue answer is revealed, the test is over.
Do not adjust the score and pretend the same object remains blind.
If the method changes, that manuscript can move into the development history but should not be reused as fresh evidence of unbiased performance.
This is the equivalent of preventing data leakage between training and test sets.
Blind Analysis Does Not Remove All Bias
The feature packet itself can be biased.
Who chose which ruling features to include?
Who defined the candidate regions?
Who decided what “Paduan-looking” means?
Blinding protects the later inference step.
It does not excuse poor feature design.
That is why calibration and out-of-sample testing must surround the blind test.
The eduKate Blind Localisation Protocol
- Select a known-origin manuscript not used to build the current scoring weights.
- Assign it a random object code.
- Remove shelfmark, institution, catalogue origin and ownership labels.
- Remove named-author information when testing physical-production geography.
- Create a frozen neutral evidence packet.
- Pre-register the candidate regions and scoring rules.
- Have the analyst rank candidates and assign confidence.
- Allow mixed and null outcomes.
- Seal the result before revealing the catalogue coordinate.
- Score the result exactly as pre-defined.
- Archive failures and do not recycle the object as a fresh blind test after model revision.
What the Blind Test Would Do to the Voynich Programme
It would make several favourite evidence classes earn their keep.
If “Paduan medical architecture” repeatedly causes blind false positives in Vienna and Germany, its geographic weight must fall.
If certain codicological bundles consistently recover known production regions, their weight should rise.
If the method often returns null on genuinely ambiguous objects, that may be a sign of discipline rather than weakness.
Research Sources
- MacCoun & Perlmutter — Blind analysis: Hide results to seek the truth.
- NCBI Bookshelf — Prediction Modeling Methodology — data leakage, training/test separation and external validation.
- Voynich and Padua — The Calibration Problem
- Voynich and Padua — The Null Coordinate
- Voynich and Padua — The Evidence Independence Problem
- Voynich Research Library
The Final Idea
When the answer is visible, evidence has a strange habit of arranging itself around the answer.
Hide the label. Freeze the evidence. Make the call. Then let the manuscript tell us whether our method deserved its confidence.