Write a few Voynich words with the meanings you propose, then press Substitute and judge.
A single word is judged by how it repeats. Several words are also judged by how they turn up near each other, compared with how the corpus uses them.
Name a meaning, and this ranks the Voynich words it would fit best.
What a substitution is tested against
The proposed meanings replace their Voynich words everywhere in the transcription (data/sample.txt; the f57v ring and other runs of spaced-out single glyphs, and line 17, are left out). The plaintext side is the dated corpus of the semantic search: Gutenberg originals and Sefaria texts composed by 1438, one stream per language.
The score
At every occurrence of a mapped word, each neighbour up to the window away on the same page is either another mapped word or something else. The corpus gives the chance of each outcome next to that plaintext word, and the chance with no association at all. The score is the log-likelihood ratio in bits, summed over all occurrences. A pair the corpus links strongly earns bits where the manuscript has it and costs bits where the manuscript never shows it.
The null
Bits alone mean little: a frequent, repetitive Voynich word piles them up either way. So the same meanings are given to random Voynich words of similar frequency (within a factor of 1.5), a few hundred times. The p-value is the share of those random assignments that score at least as well. The whole mapping is tested with every word swapped at once, and each word with only that word swapped.
The profile
Five measures that mean the same in any language: frequency, how often the word recurs within the window, how evenly it spreads over page-sized stretches, and how common its left and right neighbours are. The profile distance is not part of the score; its rank among the random pool shows whether the word is used like its meaning.
Reading the result
Trying many mappings will turn up p < 0.05 by chance about one time in twenty. Treat the p-values as a ranking, and trust a mapping when it keeps winning as you add words, not when one try comes out well. The reversed-order score is shown for comparison: a real rendering has a direction.