LangNet is a local evidence engine for reading Latin, Greek, and Sanskrit. It connects words in texts to possible headwords, morphology, dictionary witnesses, reader references, compact glosses, and source caveats.

The reliable backend contract is CLI JSON. The SvelteKit webapp adapts that contract through routes for lookup, reader texts, library/source-index browsing, word indexes, paradigms, message-of-the-day data, and translation-cache inspection.

Project Orion is the public product name: the hosted site, web UI, public copy, and learner-facing experience.

LangNet is the internal/runtime name: this repository, Python package, langnet-cli, data builders, schemas, and developer contracts. Public Orion surfaces may be powered by LangNet, but they should not require users to learn that internal name.

LangNet is useful but still in stabilization:

• word-level encounter output is the core learner surface;
• reader catalog, source-index export, /library, word-index, paradigm, and translation-cache routes exist;
• DICO, Gaffiot, and Bailly French entries can project cache-backed English learner glosses without replacing the source witness;
• reader expansion now uses source manifests, staged samples, quality flags, and checked-in source-index snapshots instead of raw corpus-count goals;
• broad passage interpretation, embeddings, and opaque generated answers remain deferred until word-level evidence and reader provenance are reliable.

The main forward work is to keep reader quality gates tight, improve selected word context in the reader, broaden compact learner glosses, and continue source structuring only where fixtures justify typed fields.

# Enter the devenv shell, then run commands normally.
devenv shell
langnet-cli --help

# Or run one-off commands through the maintained wrapper.
just cli lookup lat lupus --output json
just cli encounter san dharma all --output json
just cli translation-cache status --output json

# Inspect source-backed learner output and evidence.
just cli plan lat lupus
just triples-dump lat lupus whitakers
just cli encounter lat arma gaffiot --translation-mode cache
just cli encounter san dharma dico --translation-mode cache

# Inspect source-backed inflection tables.
just cli paradigm san putra --kind declension --gender Mas --output json
just cli paradigm san gam --kind conjugation --class 1 --output json
just cli paradigm lat amo --kind conjugation --output json
just cli paradigm grc lo/gos --kind declension --output json

Bailly and Lewis surfaces are local data/index surfaces, not replacements for the CLI encounter contract. Bailly/Gaffiot/DICO French source entries can be projected through the translation cache when exact cache rows exist or when population is explicitly requested.

• langnet-cli encounter is the learner-facing word encounter.
• langnet-cli lookup, plan, plan-exec, and triples-dump are inspection and debugging surfaces.
• word-index, reader, paradigm, translation-cache, word-of-day, and recommend-words expose reader and learning workflows.
• The SvelteKit webapp lives in webapp/ and adapts CLI-backed data through /api/search, /api/reader, /api/word-index, /api/paradigm, /api/motd, and /api/translation-cache.

External services and local data remain environment-dependent. Sanskrit Heritage, Diogenes, Whitaker's Words, CLTK data, CDSL data, and corpus/index databases must be installed or built separately for live lookup and browsing.

• docs/README.md - documentation map
• docs/GETTING_STARTED.md - setup and first commands
• docs/OUTPUT_GUIDE.md - CLI and JSON output guide
• docs/DEVELOPER.md - development workflow
• docs/ROADMAP.md - milestone roadmap
• docs/EXECUTION_PLAN.md - current active queue
• docs/READER_CLI_BEGINNER_GUIDE.md - reader corpus operation
• docs/READER_WEB_CONTRACT.md - reader/web integration contract
• webapp/README.md - SvelteKit webapp operation
• AGENTS.md - AI-agent workflow notes

This project uses multi-model AI-assisted development via OpenRouter. See AGENTS.md for specialized agent usage:

• @architect - system design and planning
• @sleuth - debugging and root cause analysis
• @coder - feature implementation and testing
• @artisan - code optimization and style
• @scribe - documentation and comments
• @auditor - code review and security

Use just lint-all, just test-fast, and focused just test ... recipes for local validation. Run recipe probes sequentially when auditing recipe health, and restart long-lived web or process-manager sessions after code changes.