# Demo: labeling before/after
## Before
Manual loop:
1. Ask for the same labeling prompt again.
2. Explain strict vs broad scope by hand.
3. Run Ark batch labeling.
4. Discover enum drift such as `POI下不展示` vs `POI 下不展示`.
5. Re-run failed samples.
6. Manually summarize distribution and representative cases.
## After
Night Gym candidate:
1. Require scope statement at the top of the prompt.
2. Normalize label enums before validation.
3. Retry failed rows with a shorter prompt and preserve JSONL cache.
4. Write `summary.md` with distribution, failures, and representative cases.
Expected effect:
| Check | Before | After |
|---|---|---|
| Scope statement | Written ad hoc | Generated from template |
| Enum drift | Manual cleanup | Normalized before validation |
| Failed rows | Manual rerun | Short-prompt retry lane |
| Distribution | Manual count | `summary.md` |