annotation/SAMPLE_MANIFEST.md

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.