Strong evidence of a structured four-phase system in the Voynich Manuscript

[Urtx13]

EDIT: Your deterministic mapping from Lunar_Angle to Phase in generate_lunar_angles_seed.py makes the Lunar_Angle a trivial predictor to Phase - is Lunar_Angle a feature fed into your ML model?


Regardless of the previous EDIT, would you mind answering these questions?

1. The sweet spot is 4 phases; any other number doesn't yield such good results?
   
2. How's the performance with a single decision tree instead of a random forest? Unless I'm doing something wrong, scores seem to be good.
   
3. You're splitting the data 75% for training and 25% for testing during the shuffle/ablation/etc. validation step, using the same seed as during training, correct?
   

Let's see:



“Lunar_Angle is a trivial predictor because it determines Phase?”

Yes — intentionally. Lunar Angle is a synthetic, deterministic mapping to a Phase. The goal is not to train a model to “guess” the phase. Instead, Phase acts as a fixed segmentation of the manuscript, consisting of 4 arbitrary but stable sections.

The actual goal is to test whether lexical entropy and topic distributions (from LDA) correlate with that segmentation. If they do, that suggests internal structure in the manuscript that aligns with this imposed 4-phase partition. That’s what we’re testing.

If Lunar_Angle didn’t predict Phase, something would be wrong with the code.



“Why 4 phases? Have you tried other numbers?”


Yes, and or entropy and topic signals collapse, or there’s no significant prediction above permutation baselines.

Only with four do we get a statistically considerable structure: validated via permutation tests, shuffling, ablation, FFT, and autocorrelation.



“How does a single decision tree perform?”


It performs well — almost too well — because it immediately splits on Lunar_Angle.
But again, the aim isn’t model optimization. It’s to test feature informativeness.
So we also test:

• Model accuracy without Lunar_Angle
  
• ablation of entropy and topics
  
• Complete shuffle tests and random permutations
  

This lets us isolate which features carry a signal, and they do.




“You’re using the same random seed across all splits?”


Of course!
We want the same train-test split across all experiments (original, permutation, shuffle, and ablation) to ensure fair and controlled comparisons.



TL;DR:

• Phase is a fixed reference segmentation, not a target of interest.
  
• Lunar_Angle trivially predicts Phase — that’s by design.
  
• The point is: does the text itself align with that structure?
  Our tests say yes — consistently, robustly, and reproducibly.
  

In short, I think you’re confusing the purpose of the experiment.

This isn’t about predicting Phase as if it’s a hidden label. Phase is a synthetic, deterministic segmentation — one that is intentionally created.

The goal is to test whether features like entropy or topic distribution align with that segmentation.
If they do (and they do, even without Lunar Angle), it suggests a latent structure in the manuscript.

So the mapping is deterministic, on purpose.
We’re not modeling truth. We’re testing alignment.