100 Wrong Verdicts a Day: The Fine Print Inside a “99% Accurate” AI SOC

**Executive summary:** This piece explains that LLM-based SOC tools can produce confident false negatives by performing probabilistic pattern completion over incomplete contexts (treating missing or failed data retrieval as ‘no evidence’), and recommends that orchestration layers enforce an explicit evidence ledger, typed nulls and gaps, confidence tied to coverage, and fail-open handoffs so missing evidence increases uncertainty rather than benign confidence.