How often each agent fails semantic integrity — the meaning could not be resolved, or the ask was outside what the agent may have. Every column sorts; there is no single score, because how much activity a count came out of is part of the answer.