A model can contain many features and still depend almost entirely on one of them.
The Padua localisation framework now looks rich:
- codicology;
- palaeography;
- image ancestry;
- astronomical parameters;
- documentary provenance;
- expected evidence;
- path complexity;
- negative controls;
- null thresholds.
But richness in the input list does not prove that all of those families matter to the output.
The Feature Ablation Test removes one evidence family at a time and asks how much the localisation result changes.
If removing a family barely moves the score, that family is not carrying much of the decision.
If removing one family flips Padua to Vienna or to null, that family is structurally important.
What Ablation Means
In machine learning and model interpretation, ablation is a simple but powerful idea: hide or remove one feature or component and measure how model performance changes. The change reveals something about that feature’s contribution to the model.
We can translate that logic into historical localisation without pretending the Voynich framework is a machine-learning classifier.
The question is still the same:
what stops working when this evidence family disappears?
Ablate Families, Not Only Individual Clues
Removing one individual plant resemblance tells us little if twenty closely related image features remain.
The stronger test removes whole evidence families.
| Ablation | What disappears | What we learn |
|---|---|---|
| Remove codicology | Ruling, quire structure, parchment-side habits, physical production clues. | Whether Padua depends on the physical object or mostly on content parallels. |
| Remove palaeography | Ductus, scribal-training bundles, hand-related regional signals. | Whether script geography materially changes the ranking. |
| Remove image ancestry | Herbal, diagrammatic and visual-genealogy evidence. | Whether resemblance is carrying more weight than admitted. |
| Remove documentary provenance | Inventories, ownership records, custody anchors. | How much the model relies on explicit historical bridges. |
| Remove local parameters | Meridian, latitude, local numerical or astronomical signals. | Whether any genuinely local parameter is decisive. |
| Remove expected-evidence penalties | Costs for predicted but missing observations. | Whether the favourite survives only because missing evidence is ignored. |
| Remove path penalties | Costs for extra handoffs and geographic stages. | Whether a complex route is being restrained by parsimony or otherwise runs free. |
Ablation Is Different From Weight Sensitivity
The Weight Sensitivity Problem asks what happens when a feature family receives more or less influence.
Ablation asks something harsher.
Remove it entirely.
Set its contribution to zero.
Then rerun the system.
That tells us whether the conclusion has multiple load-bearing supports or one hidden pillar.
The Load-Bearing Evidence Problem
Suppose the full model prefers Padua strongly.
Then:
- remove image ancestry → Padua still leads;
- remove palaeography → Padua still leads;
- remove astronomy → Padua still leads;
- remove codicology → Padua falls below Vienna.
Now we know something important.
The apparent multi-evidence Padua result is physically load-bearing on codicology.
That may be entirely appropriate.
But the public conclusion should say so.
A result should disclose which evidence family is actually holding it up.
Ablation Can Expose Decorative Evidence
Some evidence may be interesting historically but irrelevant to the final geographic decision.
For example, the presence of Paduan-linked medical authors may enrich the transmission story but contribute almost nothing once the Receiver Fallacy is properly controlled.
If removing that family leaves the physical-production ranking unchanged, the model has correctly treated it as context rather than localisation.
That is a good result.
Ablation Can Expose Hidden Double Counting
Suppose removing image ancestry causes a huge collapse in Padua score, much larger than the nominal image weight suggests.
That may mean image-derived assumptions leaked into other families.
Perhaps visual classification influenced which texts were selected.
Perhaps image ancestry indirectly altered path costs.
Perhaps several correlated image features were counted in different names.
The Evidence Independence Problem may therefore appear as an ablation anomaly.
Group Ablation Is More Informative Than Single-Feature Theatre
A large model can survive removal of one tiny feature while remaining entirely dependent on that feature’s close relatives.
So the principal units should be causal or epistemic groups:
- all botanical-image features;
- all script-training features;
- all documentary evidence;
- all local astronomical parameters;
- all codicological production features.
Only after family ablation should we drill down to individual features.
Leave-One-Coordinate-Out Is Another Useful Test
The Coordinate Stack allows a second form of ablation.
Remove an entire coordinate layer.
- no C2 text ancestry;
- no C3 image ancestry;
- no C4 scribal training;
- no C5 physical-production evidence;
- no C8 ownership evidence.
This reveals whether a physical-production conclusion is quietly dependent on evidence from a different coordinate.
If removing C2 destroys C5, the Receiver Fallacy may not have been fully controlled.
Ablate the Favourite’s Best Evidence
A particularly strong hostile test is to remove the evidence family that most helps the leading region.
If Padua still leads after its strongest family is removed, that indicates genuine distributed support.
If it collapses immediately, the result may still be valid—but it is narrower than a phrase like “many independent lines of evidence point to Padua” would imply.
Ablate the Rival’s Best Evidence Too
Fairness requires symmetry.
Remove the evidence family that most helps Vienna.
Remove the one that most helps Germany.
Remove the one that most helps Bohemia.
Compare how each candidate degrades.
A robust winner may not be the candidate with the highest baseline score; it may be the candidate whose support is least concentrated in one fragile channel.
The Ablation Matrix
| Run | Padua | Vienna | Germany | Null | Interpretation |
|---|---|---|---|---|---|
| Full model | Baseline | Baseline | Baseline | Baseline | Reference. |
| – Codicology | Δ | Δ | Δ | Δ | How load-bearing is physical production? |
| – Palaeography | Δ | Δ | Δ | Δ | How load-bearing is scribal training? |
| – Images | Δ | Δ | Δ | Δ | How much does visual ancestry drive the result? |
| – Documentary | Δ | Δ | Δ | Δ | How much does custody/provenance drive it? |
| – Local parameters | Δ | Δ | Δ | Δ | Are rare geographic parameters decisive? |
The numbers should be generated only by a future frozen model.
This article defines the test before the result.
Ablation and the Null Coordinate
Removing a family may push all candidates below threshold.
That is informative.
It tells us the removed family was not merely changing the winner; it was supplying the evidence needed to assign any coordinate at all.
Again, null is a valid result.
The eduKate Feature-Ablation Protocol
- Run the preregistered full localisation model.
- Freeze the baseline outputs.
- Remove one complete evidence family.
- Recompute all candidate scores without retuning remaining weights.
- Record score shifts, ranking changes and null transitions.
- Repeat for every evidence family.
- Repeat at Coordinate Stack level.
- Run grouped ablations for correlated families.
- Identify the most load-bearing and most decorative evidence.
- Report the ablation matrix alongside the final localisation result.
What Would Count as a Warning?
- one low-quality or subjective family causes nearly all of the result;
- removing one family changes multiple supposedly independent coordinates;
- the model requires image evidence to sustain a physical-production claim;
- several evidence families turn out to be redundant copies of one causal channel;
- the winner changes under almost every family ablation.
Research Sources
- Randomized Ablation Feature Importance — measuring feature contribution by hiding or ablating inputs.
- Voynich and Padua — The Weight Sensitivity Problem.
- Voynich and Padua — The Evidence Independence Problem.
- Voynich and Padua — The Coordinate Stack.
- Voynich Research Library.
The Final Idea
A result supported by many evidence families should be able to lose one and remain recognisable.
Remove the pillars one at a time. Then we will know whether we built a structure—or decorated one pillar.