# Validation sample manifest

- Drawn from the **test split** of `outputs/benchmark/benchmark_labelled.jsonl`
- Seed **20260810**, strata {'1hop': 50, '2hop': 50, '3plus': 20}, **120 items**
- **Excludes all 500 questions annotated in the 500-question IAA pilot** — zero overlap asserted at generation. This matters because one of the two raters worked that pilot; an item they have already labelled cannot yield an independent second judgment.
- Strata rebalanced 50/50/20 (was 40/40/40): the pilot had consumed 100 of the only 143 3-hop test questions, leaving 43 unseen. Drawing 40 would have exhausted the stratum for any future round.
- Both raters receive the **same** items in the **same** order; label
  independently and blind. **Neither rater is an author.**
- ⚠️ Disclose in the paper: one rater participated in the pilot's collaborative calibration round, so agreement partly reflects shared calibration rather than guideline clarity alone.

| complexity | operator | n |
|---|---|---|
| 1hop | before_after | 13 |
| 1hop | interval | 13 |
| 1hop | point_in_time | 12 |
| 1hop | sequence | 12 |
| 2hop | before_after | 13 |
| 2hop | interval | 13 |
| 2hop | point_in_time | 12 |
| 2hop | sequence | 12 |
| 3plus | before_after | 6 |
| 3plus | interval | 2 |
| 3plus | point_in_time | 6 |
| 3plus | sequence | 6 |

Regenerate with `python code/sample_validation_120.py` — deterministic.