How the rules work

All of Gimeltra’s knowledge lives in one file, gimeltra/gimeltra.tsv. Each row is one abjad letter; each column is a script. update.py compiles the TSV into gimeltra_data.json, which the engine loads at runtime.

The TSV

The Latn column holds the canonical romanization — the anchor every other script maps to and from.
Every other column holds that letter’s form in that script, written as a cluster of |-separated codepoints with a few sigils:

Sigil	Meaning
(bare)	the letter, used both when converting to and from this script
`>x`	`x` is written only when converting into this script
`<x`	`x` is accepted only when converting from this script
`~x`	`x` is the word-final form of the letter
`a%b`	ligature: `a` is rewritten to `b` after conversion

The special <Latn column feeds the simplify table, which folds rare Latin transcriptions onto a base letter so a missing target still resolves.

The passes

Each conversion runs three passes (see gimeltra.py):

Preprocess — apply ccmp cluster rewrites, NFD-decompose, then drop every combining mark. This is the abjad reduction.
Convert — map each character source → target, pivoting through Latin when no direct rule exists.
Postprocess — apply fina word-final forms and liga ligatures in the target script.

Regenerating the data

Edit the TSV, then rebuild the JSON:

pip install "gimeltra[update]"
python -m gimeltra.update