Please help :-) Ablation-tested phonological decoder (EVA → Elu-Sinhala)

[KamCode]

the text was probably originally written on palm leaves which is why you see very few straight lines in the script — straight lines crack the leaf.

q , k , t , p ch

The text is full of straight lines, with the most notable characters like gallows explicitly being made from long, straight vertical lines. 

the dictionary is modern sinhala, not medieval. thats a fair weakness to point out. medieval elu vocabulary would be different and i havent been able to source a proper historical elu dictionary yet. thats one reason i want a sinhala linguist involved, someone who knows the old forms.

The dictionary "sinhala_dictionary.txt", how confident are you that it's correct? The first words listed are:

"a, aa, aaa, aaaa, aaaaa, aaaaaa, aaaaaaaa, aaaah"

It doesn't give a lot of confidence. I don't mind using AI for certain tasks, but are you accepting AI output/input as fact and running with it?

hi sorry for not being clear i am saying the author originally would of made there notes from spoken word onto palm leaves, i believe it was copied from those to vellum some time later months/years to preserve them, so originally it would of been no straight lines mainly loops, the text now is a hybrid, also ill explain the dicitonary below its easier to take the explanation from ai as alot to retype in my own words, please dont mind

The dictionary: Fair challenge. The entries starting with "a, aa, aaa, aaaa, aaaah" look like junk but they're mostly not — "aaabharana" is ābharaṇa (ornament), "aaachaarya" 
  is ācārya (teacher), "aaadharana" is ādhāraṇa (support). The triple-a is a romanization artifact: "aa" represents long ā, so a word starting with ā + another a-initial       
  syllable becomes "aaa." There are about 269 entries starting with "aaa" out of 1,470,278 total — less than 0.02%. The first handful ("a, aa, aaa, aaaa, aaaaaaaa, aaaah") are 
  genuine noise — 32 entries total of repeated single characters in the entire file.                                                                                           
                                                                                                                                                                               
  The dictionary is a computationally-derived modern Sinhala romanized wordlist with full morphological inflections (beheth, behetha, behethak, behethaka, behethakata... =
  medicine + case/number suffixes). It contains real Sinhala vocabulary including medical terms (roga, aushadha, vaidya, kashaya, churna, ghrita, sneha). It is not AI-generated.

  That said — the dictionary being modern is a real limitation, as I acknowledged. And you're right to push on data quality. I should document the dictionary source explicitly
  in the paper. It currently doesn't say where it came from and it should.

thanks i will be offline till the morning but happy to answer more and any questions then thanks