Estimated

global-corpus-prior over 193 held-out anchors (51%); launch card against 2 resolved peers on 8 rows (49%); exceeds every named peer on 1 of 8 rows; optimism haircut 1.3 from cross-lab-probe; shrunk 4.1 up toward the measured corpus

Inherited

finetune of deepseek/deepseek-r1 · finetune (100%)