cookbook/environments/_06_learning_zone/README.md
Find tasks whose repeated attempts include both passes and failures. For a
binary scorer, the learning zone is exactly 0 < pass_rate < 1: neither already
mastered nor consistently failed.
basic.py — run a mixed task set and call learning_zone().select_middle_band.py — make the strict partial-pass-rate filter explicit.saturated_tasks.py — contrast saturated tasks with useful middle-band tasks.Use the learning zone to decide where more examples, prompt work, or verified dataset curation can add signal. Full-pass tasks add repetition but little new information; zero-pass tasks may be beyond the current policy.
This builds on the scorers in _03_code_scorer/,
_04_judge_scorer/, and
_05_tool_call_scorer/. Next,
_07_difficulty_calibration/ shows how to
move a task into this band.
python cookbook/environments/_06_learning_zone/basic.py
python cookbook/environments/_06_learning_zone/select_middle_band.py
python cookbook/environments/_06_learning_zone/saturated_tasks.py
Requires OPENAI_API_KEY. This is repeated verification and task selection,
not a live RL reward loop.