Tag
HyperHat's founder shares a design sketch for benchmarking whether AI agents can identify contradictory task requirements and ask clarifying questions rather than blindly executing, pairing broken tasks with satisfiable controls to prevent trivial gaming of the eval.