Model Registry

Picochat runs write many useful artifacts. The model registry turns those run folders into a single auditable table for teams deciding which checkpoint is safe to inspect, export, or deploy.

Build a Registry

From explicit run directories:

picochat registry runs/run-a runs/run-b \
  --out reports/model_registry.md \
  --json-out reports/model_registry.json

Or discover every run with summary.json under a run bank:

picochat registry --runs-dir runs \
  --out reports/model_registry.md \
  --json-out reports/model_registry.json

The registry includes:

• gate status and profile
• parameter count and planned token budget
• tokens per parameter
• eval and SFT fit rates
• held-out SFT fit when available
• external benchmark count
• honesty and preflight status
• best/resume checkpoint paths
• tokenizer path

Write a Release Card

For a single run:

picochat registry runs/h200-1b-release \
  --release-card reports/h200-1b-release-card.md

The release card is a compact artifact for handoff. It points to required evidence: preflight, honesty, eval, external benchmarks, checkpoints, tokenizer, and gate issues.

Product Rule

The registry is an index, not an approval system. A row with status approved is useful only when the linked evidence exists and the run’s model card/export matches the same checkpoint. A blocked run should stay visible so failures are not lost between experiments.