A single skill that closes the largest documented gaps between how LLMs reason and how they act—metacognition, planning, verification, and tool grounding in one control loop.
Genius here means: correct, grounded, and finished—not longer or more confident answers.
Research on LLMs and agents repeatedly finds the same bottleneck: models often know they lack information, should verify, or should abstain—but still produce plausible, confident outputs anyway. That knowing-doing gap limits reliability more than raw model size.
GENIUS addresses that with an external harness: structured gates, plan-then-execute, factored self-verification, and mandatory tool evidence—so diagnosis always becomes the next action.