orthos

Rule-first Russian proofreading engine in Rust.

This is not a regex toy. The project is already building a transparent language-model layer for Russian: tokenization, morphology, agreement, government, quantity frames, punctuation safe zones, suppressions, debug snapshots, dataset importers, and regression fixtures.

What It Already Catches

Run the strict profile to see the current research-grade surface:

cargo run -- check examples/bad.txt --rules rules --profile strict

Representative checks that already work:

Согласно новому приказа, встреча перенесена.
→ preposition + nominal-group government conflict

Два новых дом стояли рядом.
→ numeral + short nominal-group agreement conflict

Сто двадцать два новых дом стояли рядом.
→ typed compound-numeral component selects the governing numeral part

Новый важному приказу присвоили номер.
→ non-adjacent modifier/head agreement conflict inside a nominal group

Девочка пришёл.
→ subject/predicate agreement conflict

Он говорит о задачу.
→ verb valency frame: prepositional object must use the modeled case

Он помогает проектом.
→ verb valency frame: direct-object seed expects another case

Он хочет разсказать и потом расбить лёд.
→ word-formation-backed prefix-final з/с assimilation

Из за дождя кое кто опоздал. Я то понял. Ну ка проверь.
→ model-backed hyphenation for compound prepositions, кое-, -то, -ка

Он ушёл потому что устал. Конечно это риск.
→ punctuation with clause-marker and introductory-phrase safe-zone logic

Мaма сказала: я незнаю, но надо учится.
→ mixed alphabet, common не-confusable, and -тся/-ться heuristic

Пол лимона осталось.
→ conservative пол- hyphenation before л/vowel/proper-name contexts

The engine also keeps negative fixtures for false-positive risk. For example, corrected targets from LORuGEC, RLC, and RuSpellGold are stored as regression fixtures so model-backed grammar rules do not keep firing on normal sentences like В один из летних вечеров..., От него целый год..., or к потере слуха и зрения.

Philosophy

The long-term product direction is a serious Russian grammar/proofreading checker with minimal ML and maximal explicit linguistic structure. Regex is allowed only for surface phenomena. Grammar rules should be backed by reusable models:

• typed morphology and feature unification;
• preposition and verb government frames;
• nominal-group and quantity models;
• clause boundaries and syntactic safe zones;
• debug snapshots with machine-readable proofs;
• dataset-based regression tests.

Install on Arch Linux

sudo pacman -S --needed rust cargo jq parallel python-pytest

Run

cargo run -- check examples/bad.txt --rules rules
cargo run -- check examples/bad.txt --rules rules --profile strict
cargo run -- check examples/bad.txt --rules rules --format json | jq

Use a custom morphology TSV:

cargo run -- check input.txt --rules rules --morph-lexicon data/lexicon/demo_morph.tsv

Inspect the debug layer:

cargo run -- debug input.txt --rules rules --profile strict | jq

Validate the rule corpus:

cargo run -- validate --rules rules

Run embedded rule examples as executable specs:

cargo run -- test-examples --rules rules

Batch check with GNU Parallel:

find texts -type f -name '*.txt' -print0 | \
  parallel -0 'cargo run --quiet -- check {} --rules rules --format json > {.}.lint.json'

Profiles and Filtering

Default checks run conservative implemented rules only. strict enables all implemented rules, including style, heuristic, and advanced demo/model-backed checks.

cargo run -- check examples/bad.txt --rules rules --profile default
cargo run -- check examples/bad.txt --rules rules --profile strict
cargo run -- check examples/bad.txt --rules rules --profile typography-only
cargo run -- list-rules --rules rules --all --status research

Targeted filters:

cargo run -- check examples/bad.txt --rules rules --domain punctuation --severity error
cargo run -- check examples/bad.txt --rules rules --rule-id ru.punctuation.no_space_before_mark
cargo run -- check examples/bad.txt --rules rules --exclude-rule ru.typography.hyphen_instead_of_dash

Suppressions are off by default. Enable inline directives explicitly:

cargo run -- check input.txt --rules rules --allow-inline-suppressions
cargo run -- check input.txt --rules rules --suppress-rule ru.punctuation.no_space_before_mark

See docs/CLI_USAGE.md for the JSON contract, suppression syntax, timing mode, and batch examples.

Evaluation Datasets

Local-only importers currently support:

• RLC annotated / RLC-GEC;
• LORuGEC;
• RuSpellGold.

The repository does not vendor full raw corpora by default. Use local archives from your machine:

python scripts/eval/run_benchmark.py \
  --dataset lorugec \
  --archive "$HOME/Downloads/LORuGEC-main.zip" \
  --checker-bin target/debug/orthos \
  --rules rules \
  --profile strict \
  --limit 100

The checked-in fixture testdata/fixtures/eval/gec_dataset_regressions.jsonl keeps a tiny, stable slice of real dataset-derived false-positive cases for normal development.

Full Morphology Cache

The repository includes data/lexicon/demo_morph.tsv, a small built-in lexicon used by tests and examples. For broader morphology-backed checks, download the OpenCorpora-derived cache from the GitHub release:

mkdir -p data/lexicon
curl -L -o data/lexicon/opencorpora.bincache.zst \
  https://github.com/xa9e/orthos/releases/download/morph-cache-v1/opencorpora.bincache.zst
curl -L -o data/lexicon/opencorpora.bincache.idx.zst \
  https://github.com/xa9e/orthos/releases/download/morph-cache-v1/opencorpora.bincache.idx.zst

zstd -d data/lexicon/opencorpora.bincache.zst -o data/lexicon/opencorpora.bincache
zstd -d data/lexicon/opencorpora.bincache.idx.zst -o data/lexicon/opencorpora.bincache.idx

After that, orthos automatically uses data/lexicon/opencorpora.bincache when no --morph-lexicon is passed.

The cache is derived from the OpenCorpora morphological dictionary. OpenCorpora data is distributed under CC BY-SA 3.0; the project code is licensed under AGPL-3.0-or-later.

Test

cargo fmt --check
cargo check
cargo test
cargo run -- validate --rules rules
cargo run -- test-examples --rules rules
python3 scripts/dev/project_health.py
python3 -m pytest -q tests/test_importers.py tests/dev_tools/test_project_health.py
git diff --check

Current Detector Families

• whitespace, punctuation spacing, sentence-final punctuation, lowercase starts;
• repeated words and repeated punctuation;
• mixed Cyrillic/Latin words;
• common spelling and confusable-token seeds;
• clitic and compound-preposition hyphenation;
• не + verb/gerund/participle spacing;
• -тся/-ться context heuristic;
• пол- hyphenation in conservative contexts;
• prefix-final з/с assimilation through a seed word-formation model;
• missing comma before frequent subordinators with safe-zone suppression;
• introductory phrase comma at sentence start;
• coordination comma slots for homogeneous modifier series;
• adjective/noun and nominal-group modifier agreement;
• subject/predicate agreement;
• preposition case government;
• preposition + nominal-group government;
• numeral/noun and numeral + nominal-group quantity agreement;
• compound and typed-component compound numeral agreement;
• verb government from data-backed valency seeds;
• phrase-map style, pleonasm, and government seeds;
• debug proof generation for model-backed diagnostics.

Development Notes

Read IDEA.md first. The project direction is rule-first and debug-first: every serious grammar rule should become a small inspectable language model, not a one-off detector.

Useful docs:

• docs/DATASETS.md
• docs/CLI_USAGE.md
• docs/ARCHITECTURE.md
• docs/language-model/
• docs/debug-layer.md

License

The project is licensed under the GNU Affero General Public License, version 3 or any later version (AGPL-3.0-or-later). See LICENSE for the full text. External datasets keep their own licenses and are never vendored into this repository; see docs/DATASETS.md.

Git Bundle Workflow

Create a portable repository bundle:

git bundle create orthos.bundle --all
git bundle verify orthos.bundle

Restore it:

git clone orthos.bundle orthos
cd orthos

For a source-only archive without Git history:

git archive --format=tar HEAD | xz -T0 > orthos-source.tar.xz