# Evaluator-only combined episodes

These JSONL files combine each episode's inputs and target values in one record
for the scorer and offline statistical baselines:

- `MODEL_DEV67_EPISODES_V1_WITH_TARGETS.jsonl`
- `MODEL_TEST211_EPISODES_V1_WITH_TARGETS.jsonl`

Use `inputs/` for agent runs. Keep `targets/` and these combined records in the
evaluation environment: they contain ground-truth values such as
`future_counts` and `future_share` that must not enter model prompts.