Resample by hand
21 of 30 graded outputs are correct (emerald), 9 are incorrect (rose): point estimate accuracy 0.70. Each resample draws 30 items WITH REPLACEMENT and recomputes accuracy.
Distribution of replicate accuracies so far
Width vs sample size
Tile the same 21/30 set to a bigger n and bootstrap it again (3,000 resamples). The interval should stay centered near 0.70 and shrink as n grows.
At n=30 the companion script's full 10,000-replicate bootstrap reads 95% CI = [0.5333, 0.8667], width 0.3333. Tiling to n=120 halves the width to about 0.1667; tiling to n=480 halves it again to about 0.0833. This panel uses 3,000 live resamples per setting, so its numbers will land close to, but not always bit-identical to, that fixed-seed run.