| 2026-07-17T04:01:37-07:00 === correctness A/B start === |
| 2026-07-17T04:01:37-07:00 scoring top-128 -> outputs/teacher_trajectories/dolci_math_nogold_matched_top128 (GPU 0) |
| 2026-07-17T04:21:00-07:00 scoring complete: "total_records": 9918, |
| 2026-07-17T04:21:00-07:00 healing glean_keep25_nogold_s1224 on GPU 1 |
| 2026-07-17T04:21:00-07:00 healing glean_keep25_nogold_s1225 on GPU 2 |
| 2026-07-17T04:38:30-07:00 === correctness A/B start === |
| 2026-07-17T04:38:30-07:00 scoring complete: "total_records": 9918, |
| 2026-07-17T04:38:30-07:00 healing glean_keep25_nogold_s1226 on GPU 2 |
| 2026-07-17T04:38:30-07:00 healing glean_keep25_nogold_s1224 on GPU 0 |
| 2026-07-17T04:38:30-07:00 healing glean_keep25_nogold_s1225 on GPU 1 |
| 2026-07-17T04:41:56-07:00 === correctness A/B start === |
| 2026-07-17T04:41:56-07:00 scoring complete: "total_records": 9918, |
| 2026-07-17T04:41:56-07:00 healing glean_keep25_nogold_s1224 on GPU 0 |
| 2026-07-17T04:41:56-07:00 healing glean_keep25_nogold_s1225 on GPU 1 |
| 2026-07-17T05:33:14-07:00 eval glean_keep25_nogold_s1224 |
| 2026-07-17T05:35:05-07:00 eval glean_keep25_nogold_s1225 |
| 2026-07-17T05:35:37-07:00 glean_keep25_nogold_s1224 done -> 0.41091736163760423 |
| 2026-07-17T05:35:37-07:00 healing glean_keep25_nogold_s1226 on GPU 0 |
| 2026-07-17T05:37:25-07:00 glean_keep25_nogold_s1225 done -> 0.4116755117513268 |
| 2026-07-17T05:37:25-07:00 lane GPU1 done |
| 2026-07-17T06:28:01-07:00 eval glean_keep25_nogold_s1226 |
| 2026-07-17T06:30:25-07:00 glean_keep25_nogold_s1226 done -> 0.42532221379833207 |
| 2026-07-17T06:30:25-07:00 lane GPU0 done |
| 2026-07-17T06:30:25-07:00 === correctness A/B COMPLETE === |
| seed | curated(correct) | nogold(unfiltered) | delta |
| 1224 | 0.4131918119787718 | 0.41091736163760423 | -0.2 |
| 1225 | 0.4245640636846095 | 0.4116755117513268 | -1.3 |
| 1226 | 0.422289613343442 | 0.42532221379833207 | +0.3 |
|
|
| MEAN curated=0.4200 nogold=0.4160 delta=-0.40 pts |
| Verdict: |delta| small -> correctness barely matters (general column can skip gold-curation). |
|
|