Factory Packaging

A shipped specialist needs a small, inspectable package. The package metadata is committed; the large model or adapter artifact usually stays under ~/.cache/posttrainllm/models/ or another local/cache path.

Package Layout

specialists/<specialist-id>/
  artifact-manifest.json
  model_card.md
  eval_report.json
  tinygpt.lock.json
  prompt.md
  report.md

Use specialists/qwen3-4b-file-ops-distilled/ as the current pattern.

Required Model Card Fields

• what the specialist is for
• what it must not be used for
• base model
• artifact path
• training method
• primary eval result
• regression/breadth result
• known limits
• routing requirements if any

Required Lock Fields

• base model id and revision
• architecture
• precision
• artifact path
• tokenizer/template assumptions
• eval suite ids
• package date

The lifecycle manifest is the runtime authority for artifact kind, pinned base and tokenizer, bounded production history, compatible runtimes, and legal next actions. The lock remains package metadata and may repeat selected fields for human inspection, but must not contradict the manifest. See artifact-lifecycle.md.

Ship Criteria

Create a specialist package only when decision.json says ship.

For a newly produced package, artifact.json.lifecycle_manifest must name the adjacent lifecycle sidecar and factory-run publish-check must validate it. Legacy packages without a sidecar remain inspectable with an explicit warning; do not invent missing identity to silence that warning.

Do not package:

• exploratory runs
• failed candidates
• candidates without baseline comparison
• candidates that require a moving eval
• candidates whose artifact path cannot be resolved