RECORD № 0001·filed 2026-07-27·revision 1 · re-judged when better judges exist
forget · best config
gpt-4o observer · top-k 84 · obs 60 · reader v3
REPRODUCED
runs 81.8 · 82.4 · 81.8 → mean 82.0 · run σ 0.35 · sampling ±3.4 pp (95% CI) · published 81.8
manifest → Rekor № pending
transcriptssealed · 10 years · hash published at filing — §8 of the method
revisionsr1 (current) — original filing. re-judgings will chain here, strikethrough preserved.
source runsresearch/longmemeval/runs/ in the forget repository — per-question outputs, configs, judge templates.