Chertma

Type plain ASCII on an ordinary keyboard. Get correct Uzbek — new Latin, old Latin or Cyrillic.

input:   togri gap, ozbekcha yozish oson
output:  töğri gap, özbekça yoziş oson          (new Latin)
         toʻgʻri gap, oʻzbekcha yozish oson     (old Latin)
         тўғри гап, ўзбекча ёзиш осон           (Cyrillic)

input:   хозир келаман, рахмат                  (typed on a Russian layout)
output:  hozir kelaman, rahmat

input:   kelaslar qisela balu kettik
output:  kelaslar qisela balu kettik            (dialect is not touched)

Every line above is real output of engine/ with data/lexicon-lite.bin.

Try it: chertma.maqsudjon.com — works offline after the first visit, sends nothing anywhere.

---

Why

Uzbekistan is moving its Latin alphabet from sh ch oʻ gʻ to ş ç ö ğ. For years people will write in three scripts, and no phone keyboard has the new letters. Chertma restores what a keyboard destroyed — and nothing else.

The one rule: a correction is allowed only if the input and the corrected word have the same skeleton — the word with every distinction a keyboard could lose (ş/sh/s/w, ö/oʻ/o, ğ/gʻ/g, apostrophes, Cyrillic vs Latin) folded away. If no real word shares the skeleton, the input comes back byte for byte. We correct orthography. We never touch morphology, dialect, or voice. Details: docs/SPEC.md.

Engine

Plain ES modules, zero dependencies, no DOM, no network access.

import { Chertma } from './engine/index.js';

const c = new Chertma({ script: 'new' });        // 'new' | 'old' | 'cyrillic'
await c.load(bytesOfLexiconLiteBin);             // or load(url) with a `loader` option

c.autocorrect('togri gap');                      // 'töğri gap'
c.suggest('ozbe', null);                         // [{ word: 'özbekiston', score, source: 'lexicon' }, …]
c.convert('Тошкент', 'cyrillic', 'new');         // 'Toşkent'
c.learn('qisela');                               // on-device only
c.export();                                      // user model for backup
c.stats();                                       // { lexiconSize, memoryBytes, loadMs }

Measured

	Budget	Measured
engine/ gzipped	≤ 30 KB	12.0 KB
lexicon-lite.bin gzipped (50 000 words, 200 000 bigrams)	≤ 2 MB	0.74 MB
lexicon-full.bin gzipped (400 000 words, 2 000 000 bigrams)	≤ 8 MB	7.34 MB
load() lite	< 400 ms	24 ms
suggest() p95, cold V8, 10 000 real prefixes	< 5 ms	0.8 ms
Heap growth with lite loaded	≤ 25 MB	0.8 MB with forced GC (engine's own estimate 4 MB)

Lexicon built from 1.75 billion word tokens of tahrirchi/uz-books-v2 (MIT) and tahrirchi/uz-crawl (Apache-2.0); see data/SOURCES.md and docs/CHECKPOINT-2.md.

Quality, measured on an unfiltered sample of real text (npm test, golden suite; full run in docs/TEST-OUTPUT.txt):

• Latin text with apostrophes removed: 695 corrections, 1 wrong (99.9 % precision), 290 missed.
• Cyrillic typed on a Russian layout: 471 corrections, 14 wrong (97.1 %), 686 missed.
• Round trip new Latin → Cyrillic → new Latin holds for every lexicon word except a documented class (open question Q2b); uz-books-v2's own Latin and Cyrillic splits agree with our rules on 99.83 % of 29.9 million aligned words.

Known limits

Chertma is conservative on purpose — a missed correction is a minor loss, a wrong one is the failure it exists to avoid.

• A word that is already a real word is never rewritten. sosib pisib togri becomes şoşib pisib töğri: pisib ("sneaking up") is a real word, so it stays, even though şoşib pişib is the idiom. Open question Q19.
• Words the lexicon does not know stay exactly as typed, in every output script.
• Names and foreign words in capitals mid-sentence are only corrected to known proper nouns (Sunday, Elon stay as they are).

Every open question and every default taken is in docs/OPEN-QUESTIONS.md.

Repository

engine/      the product — plain ES modules
web/         the demo site and PWA (tools/build_site.py → _site/ → gh-pages)
tests/       node:test suites, golden file, fixtures        npm test
tools/       Python corpus pipeline and build scripts (build time only)
data/        lexicons, sources, build report
bench/       uz-alphabet-bench draft — not published, awaiting review
docs/        SPEC, open questions, status

Licence

The engine and site code have no licence chosen yet. Data licences are in data/SOURCES.md; the JetBrains Mono font is under the SIL Open Font License (web/fonts/OFL.txt).

Maqsudjon Polatov · maqsudjon.com