this is completely true but kinda pedantic in this context tbh. @GregKamradt has mentioned many times that task-specific harnesses are what's really discouraged here. this makes perfect sense since they leak the harness designer's own intelligence and hypotheses about how to solve this particular problem category into the model's efforts by implicitly constraining it to certain directions in a vast solution space. any "harness" that is general enough to be behind the official customer-facing API of a frontier lab is still fair game (yeah even including the token sampler).