Automated usually proposes candidate changes and verifies each one against a fixed task set. HarnessLens instead makes verification budget-aware: it co-explores task space and configurable components, derives candidate modifications from execution trajectories, and only evaluates a candidate on tasks related to the behavior it affects, using an attributable-evidence gate.
Across three and four benchmarks, this approach improves average held-out performance by 7.6-13.6% while consuming substantially less evaluation budget than competing baselines. The result is interesting because it addresses a known failure: fixed or random batches waste rollouts and let aggregate scores hide specific regressions. Behavior-aware verification makes evolution more reliable and sample-efficient under constrained interaction budgets.