Recipes #

Recipes are small end-to-end agent adaptation problems. They share training, promotion, native serving, smoke, and Eve eval machinery while owning their own data generation, instructions, target contract, and local checkpoint pointer.

Included recipes #

• capability-placement teaches Eve's project ontology.
• context-selection teaches selective retrieval from candidate context, including freshness, scope, multi-snippet joins, and insufficient evidence.
• action-authorization teaches a compact custody boundary: execute, draft, ask, or refuse.

Adding one #

1. Add training/recipes/<recipe_name>.py with a Recipe value named RECIPE.
2. Generate deterministic Case rows for train, validation, and hidden test.
3. Use unique slugs across every split. Do not tune against hidden-test outputs.
4. Implement strict target validation. Canonical key order is part of the contract.
5. Add recipes/<recipe-id>/instructions.md with the exact output contract.
6. Register the recipe in training/recipes/registry.py and agent/lib/recipe.ts.
7. Add assertions that cover the recipe's meaningful edge cases.
8. Run pnpm data:write, pnpm check, and the recipe dry run.
9. Measure baseline validation before training.
10. Train and promote only through the shared gate. Record model evidence separately from code completeness.

Keep examples synthetic and public-safe. A recipe should demonstrate a real difference in agent behavior, not merely add new nouns to an existing classifier.