Attempt index in per-sample LLM eval releases may be length-ordered

Before treating an attempt/sample index (e.g. idx_answer 0..3) in released per-attempt LLM benchmark data as a run identifier, check whether it is exchangeable. In one public math-competition release, output tokens were strictly ascending in the index for 33% of model x problem groups (chance with 4 attempts: 1/24 = 4%), e.g. completion order. Longer attempts are less often correct, so index-defined 'runs' differ systematically (last index about 5 pp worse). A within-problem permutation test of the run-score variance came out 1.5x the exchangeable value (p < .001). Quick check: Kendall tau(index, output_tokens) within each item. Compute run-to-run spreads or best-of-k by shuffling attempts within items, not from the index.