What Does a Harness Repair? Four Preregistered Studies Around a Longevity-Tuned Language Model

원문 초록

A harness search around a frozen longevity-tuned 1.7B model reported a rescue. An audit traced its gain to rows the start left unparsed and found that the caller sent reference answers as assistant turns; replication through the same pipeline missed this defect. We preregistered four studies and report every endpoint. In the first two, thinking off raised test accuracy over capped thinking in 7 of 12 cells, mainly by parsing rows the cap left unparsed, and was not meaningfully worse than a rescue in 7 of 9 small-model test cells. Replication reproduced an injected defect's distortion in 9 of 20 evaluable cases. At 14B, two GEPA harnesses beat thinking off by 34 test points but were not detectably above the strongest constant label. In the third study, GEPA beat the constant on FPB (confirmatory), thinking off was not meaningfully worse than any BoolQ harness, and 3 of 4 MMLU-Pro searches kept their seed; the fourth comparison was inconclusive on test. On exploratory Autocast, harnesses were within 0.7 points of the constant and gave one label in 93.1% to 96.8% of parsed answers on test. On the prediction tasks (14B mortality, Autocast) the harnesses ended near the constant label. The fourth study varied search-set skew and objective on CREAK and resolved Manifold questions. All its results are descriptive because render hashes were not committed before the first replication call (D001). Minority-recall estimates were lower at 80:20 than at 50:50 in all four test cells. The balanced objective and re-selection from each search's pool had higher minority- and lower majority-recall estimates in all four cells. Five of their eight minority-recall intervals included zero, as did all eight balanced-accuracy contrast intervals. This does not show that skew explains the earlier collapse.