Skip to main content

Deploy to Production

You proved it works in the Arena. Now deploy it.

The Arena is for R&D — building, benchmarking, and comparing translation methods. Production deployment happens through champollion, the developer-facing translation CLI. They connect through a shared plugin format.

The Deployment Path​

1. Export Your Method as a Plugin​

Create a method.json manifest that packages your benchmark results:

{
  "name": "crk-coached-v3",
  "type": "llm-coached",
  "version": "3.0.0",
  "description": "Coached LLM translation for Plains Cree",
  "locales": ["crk"],
  "config": {
    "model": "google/gemini-2.5-flash",
    "temperature": 0.3
  },
  "benchmarks": {
    "crk": {
      "composite_score": 0.67,
      "fst_acceptance": 0.82,
      "corpus_size": 150
    }
  }
}

Include any coaching data (grammar rules, dictionaries) alongside the manifest.

2. Install in Champollion​

champollion plugin install ./my-method-plugin/

3. Configure Your Pair​

champollion.config.json
{
  "pairs": {
    "en-crk": { "method": "plugin", "plugin": "crk-coached-v3" }
  }
}

4. Translate Real Content​

npx champollion sync

Your benchmarked method is now producing real translations in production.

For Indigenous Languages​

Methods serving Indigenous language communities require community consent before production deployment. The OCAP principles (Ownership, Control, Access, Possession) govern how translation methods are developed, evaluated, and deployed.

A method that reaches Deployable tier (0.70+) does not automatically deploy — it deploys if and when the language community's governance body gives consent.

See Data Sovereignty and Ownership Transfer for the full governance framework.

See Also​