Telling an agent the knobs are architecture beat anonymous variables by 12.3%, until a critic loop erased it.

arXiv 2609.19387 hands the same agent the same 15-dimensional accelerator space twice: once as named architectural knobs with simulator counters, once as anonymous variables on [0,1], with evaluator and reachable optima identical. On a nine-kernel FP16 GEMM basket the informed agent beat a modeled H200 by 5.4% and its blind counterpart by 12.3%, using 70.1% fewer simulator calls. A critic loop recovered most of the blind agent's gap and bought the informed one nothing. arXiv Domain knowledge and structured critique act as substitutes. Authors flag five to six runs per condition on one modeled accelerator as preliminary.