WHALE (Stanford/KAIST/KRAFTON) Gains Up to 24pp by Co-Evolving Agent Weights and Harness Code

Formalizes and quantifies the harness-weight coupling problem: freezing either component creates a training bottleneck, and alternating fine-tuning with harness search closes that gap by 4–24pp—adding institutional weight (Chelsea Finn, Stanford) to the emerging consensus that scaffolding matters as much as the model.