Frontier and Center: Who evaluates the evaluations?

Google Data Cloud's frontier AI team introduced Discovery Bench, a meta-benchmark using iterative surprisal-based query refinement (iSQR) to modulate evaluation difficulty. The approach generates high, medium, and low ambiguity query variations grounded in bits. Testing an F1 agent on Gemini 3.1 Pro against KramaBench revealed performance cliffs and a sweet spot that static benchmarks miss.