Runs that trace the gap between this report's register result (30 of 63 Lincoln pieces to Hay under the
2019 design) and the blind rerun's (19 of 63). Produced by diag_gap.py; file names give the pool variant
(mine = this report's grieve-design pools; nodebates = mine without the 19 Lapsley debate-volume items;
kept = mine restricted to the rerun's kept ids; theirs = the rerun's pools.json exactly), the sample size as
a fraction of the smaller pool (f095 = 0.95, f100 = 1.0), the number of random sequences, the seed, an
explicit sample size in words where one was forced (_t242957, _t159173), and _test4 for runs on the rerun's
own 97 Lincoln and 77 Hay pieces instead of this report's 120 register pieces.
*_summary.json: Lincoln pieces wrong by the character 4-10 vote and by all 17, Hay pieces wrong, per-length
accuracy, the letter's votes and margins. The full per-piece scores are in the matching .json without _summary.