Export and serve a policy

A trained policy leaves the experiment as a bundle — one movable directory: manifest beside checkpoint — and comes back as an inference run that writes the same trace training writes, so it is scored with no special case.

Name the policy, export during the run, serve from the bundle:

algorithm:
  name: discrete_hill_climbing
  params: {seed: 1}
  policy:
    id: skyfall_crl.train.backends.hill_climbing:linear_menu_policy
    kwargs: {n_actions: 2, seed: 1}

skyfall-crl run --config experiment.yaml --export bundles/my-policy
skyfall-crl serve --bundle bundles/my-policy --steps 200 --seed 7 --traces served.jsonl
skyfall-crl eval --traces served.jsonl

Tutorial L5 runs this whole flow, gate-verified, including the served table. The parts that go wrong in practice:

• The policy must be named in the document. A backend's internal default policy cannot be recorded, and a bundle without a policy reference cannot be served.
• A custom array policy needs load_weights(path). A bundle whose weights the rebuilt policy cannot take is refused, never restored empty — a partially restored policy acts like something untrained while every log line reports success.
• --seed on any serve you will score; unseeded inference is legitimately different every time. --auto-reset for environments with genuine terminal states, same as training.
• The recorded schedule is replayed by default, so served traces segment like training traces; --stationary declines it, and --env/--env-kwargs deliberately override the environment when serving against a different world is what you asked for.
• A matrix export (run --matrix --export DIR) gives every cell its own bundle subdirectory; --export-external records a path instead of copying weights, and the bundle then only works where that path resolves.