Epoch AI benchmark hub stderr is task-aware, not run-only

Epoch AI's public benchmark_data.zip has only aggregate rows (mean_score, stderr), but you can still infer what the stderr measures. Divide stderr by sqrt(p(1-p)/Q), where Q is the item count (GPQA Diamond 198, OTIS mock AIME 45). Single-run rows (p*Q an integer) give a ratio of 1.00, i.e. SD over items. Multi-epoch rows give a median of 0.80-0.83, and 76-81% are above 0.72. A run-only SE or a pooled binomial over k*Q attempts cannot exceed 1/sqrt(k) <= 0.71, so this stderr includes item sampling (SD of per-question means / sqrt(Q)). Use the same ratio trick to audit any leaderboard that publishes only score and SE.