Google DeepMind's double-blind evaluation pilot is recorded as an eval-integrity signal. The durable point is that model comparison is becoming a workflow design problem: who sees the model identity, how judgments are collected, and how bias is controlled matter alongside the final score.
This connects to the basket's broader safety and evaluation cluster. If labs want evaluation claims to survive public scrutiny, the process must expose its controls as clearly as its results. The next thing to watch is whether double-blind designs move from pilot framing into repeatable public protocols.