Zagora Discovery Lab cracks LoRA transfer

This is a strong demo of proxy search doing real work, but the bigger takeaway is how ordinary the winning recipe looks once the noise settles out. The process matters as much as the result: cheap exploration, stricter confirmation, then one honest cross-scale check.

• –The 4.14% discovery gain compressing to 1.48% after 3-seed confirmation is exactly the kind of variance haircut you want from an autonomous search loop
• –The rebound to 3.35% on 70B suggests the 8B proxy was useful, not merely overfit to itself
• –Rank 4 on all 7 modules beating rank 8 on only q/v is a nice reminder that adapter coverage can matter more than raw rank
• –The single 70B run is still proof-of-concept, not a universal law; the repo validates a recipe, not a theorem
• –Because the win is hyperparameter-only, it should transfer across distributed fine-tuning stacks, not just Zagora