Saved Baselines

Persist rollout evidence as plain JSON so a later run can be compared with the same task-level history and fingerprints. Saved artifacts include full prompts and responses and should be handled as sensitive evaluation data.

Files

• basic.py — run an environment and save its result as a baseline.
• reload_baseline.py — reload the artifact and verify its summary survived the round trip.
• async_save_load.py — use the async rollout, save, and load twins.

When to use

Use this when the baseline and candidate cannot run in the same process, or when CI needs a reviewed reference artifact. Continue to _14_environment_diff/ to compare compatible results task by task.

A baseline is evidence from a particular environment and policy, not a promise that future tasks or prompt edits remain comparable.

Run

python cookbook/environments/_13_saved_baselines/basic.py
python cookbook/environments/_13_saved_baselines/reload_baseline.py
python cookbook/environments/_13_saved_baselines/async_save_load.py

Requires OPENAI_API_KEY. Every example uses gpt-5.5 through OpenAIResponses.