Google's RRSI paper starts from an uncomfortable observation. Harness evolution works a lot like training a model on your eval set: an agent's harness gets edited task by task, the benchmark score climbs, and nobody checks what happens on a benchmark the loop never saw. When