24-02-2026, 11:35 PM
Thanks for sharing this. I actually ran the Python against the ZL3b transcription as suggested, and it does produce some nice-looking outputs. The transition heatmap is genuinely interesting to look at, but I think there are some issues worth flagging before we call this definitive.
The 100% hardware compliance result is unfortunately a ceiling effect. The atoms that the lock table blocks Ie. j, v, x, b, g, q barely exist in the manuscript anyway. They collectively account for well under 1% of tokens. You'd get 100% compliance from literally any EVA text, including random strings drawn from EVA frequency distributions. So the test doesn't actually distinguish your model from "no model at all."
The proximity bias in the inertia curve appears correct and worth exploring, but the rotor order was chosen to place high-frequency glyphs adjacent to each other. Measuring short-jump preference against that same ordering is circular and you've just arranged the wheel to fit the data, then used the data to validate the wheel. A proper test would need the rotor order to be derived independently of the frequency statistics, or at minimum tested against a frequency-weighted null model rather than a uniform random baseline.
The lock table has the same problem. Without independent evidence telling us which modes block which atoms, the assignments look reverse-engineered from what doesn't co-occur in the text. That makes them descriptions, not predictions.
Also worth noting: the cleaning function drops all tokens of 2 characters or fewer (or, ar, al, dy, ok...), which silently removes a meaningful chunk of the corpus.
None of this means a volvella wasn't involved as it's a historically plausible device for the period, albeit as we discussed, only a paper one. But this analysis as it stands doesn't provide evidence that it was. What would make it convincing: derive the rotor order and lock constraints from an independent source (a surviving volvella design, a contemporary description), THEN test against the manuscript. If it still fits, you've got something.
The 100% hardware compliance result is unfortunately a ceiling effect. The atoms that the lock table blocks Ie. j, v, x, b, g, q barely exist in the manuscript anyway. They collectively account for well under 1% of tokens. You'd get 100% compliance from literally any EVA text, including random strings drawn from EVA frequency distributions. So the test doesn't actually distinguish your model from "no model at all."
The proximity bias in the inertia curve appears correct and worth exploring, but the rotor order was chosen to place high-frequency glyphs adjacent to each other. Measuring short-jump preference against that same ordering is circular and you've just arranged the wheel to fit the data, then used the data to validate the wheel. A proper test would need the rotor order to be derived independently of the frequency statistics, or at minimum tested against a frequency-weighted null model rather than a uniform random baseline.
The lock table has the same problem. Without independent evidence telling us which modes block which atoms, the assignments look reverse-engineered from what doesn't co-occur in the text. That makes them descriptions, not predictions.
Also worth noting: the cleaning function drops all tokens of 2 characters or fewer (or, ar, al, dy, ok...), which silently removes a meaningful chunk of the corpus.
None of this means a volvella wasn't involved as it's a historically plausible device for the period, albeit as we discussed, only a paper one. But this analysis as it stands doesn't provide evidence that it was. What would make it convincing: derive the rotor order and lock constraints from an independent source (a surviving volvella design, a contemporary description), THEN test against the manuscript. If it still fits, you've got something.
![[+]](https://www.voynich.ninja/images/collapse_collapsed.png)