Training Recipes

v0.77 adds src/closed_world_lm/training_recipe.py and two required run artifacts for self-improvement answer cycles and transformer answer-training runs:

• training_recipe.json
• constraint_first_promotion.json

The recipe records how to reproduce a run. The constraint-first report records whether a run is even allowed to consider quality metrics for promotion.

Recipe Contents

training_recipe.json records:

• recipe id and version;
• component and run id;
• model configuration;
• tokenizer provenance;
• data sources;
• objective settings;
• optimizer settings;
• replay status;
• planned artifacts;
• required gates;
• rerun surface.

This keeps recipe identity out of hidden argparse memory. A future screen should be reconstructable from the recipe artifact and the admitted project state.

Constraint-First Promotion

constraint_first_promotion.json separates constraints from quality checks.

Constraints must pass before quality metrics are considered. For transformer runs, those constraints include closed-world verifier approval, closed-world training data, no pretrained weights, no pretrained tokenizer, no external embeddings, direct-answer evidence, branch-context coverage, branch diversity, and target-coverage preservation.

Only after those constraints pass can exact direct-answer quality checks affect promotion. Loss, NLL, rank, and top-k movement remain advisory until the constraint report says quality metrics are eligible.

Status

Self-improvement runs can still promote through their existing exact-eval discipline, now with a constraint-first report attached. Transformer runs now have a promotion gate, but current evidence is still expected to reject until the model satisfies the branch-context, diversity, coverage, and exact-answer quality checks.