311 operations want anomaly detection on civic AI prompts. Which monitoring foundation best enables that detection?

Baselining normal prompt patterns so anomaly detection can flag campaign-like abuse is the monitoring foundation that enables 311 civic AI defenses: without a model of ordinary resident ask patterns, every prompt looks equally suspicious or equally benign. The correct practice works by characterizing volume, topics, rates, sources, and sequence shapes under normal operations, then scoring deviations that resemble coordinated scraping, injection campaigns, or flood abuse. Aggregating ward prompt volumes so anomaly detection flags monthly demand changes fails as an anomaly-detection foundation because monthly aggregates are too coarse and delayed rather than behavioral baselines, so they cannot score deviations or flag campaign-like abuse in near real time. Retaining prompt transcripts so anomaly detection can review past abuse incidents fails as an anomaly-detection foundation because it is reactive review after the fact and does not establish the normal behavioral baselines needed to flag campaign-like abuse as it occurs. Tracking prompt inference latency so anomaly detection can flag peak slowdowns addresses infrastructure performance, not prompt-content anomalies, so it cannot serve as the monitoring foundation for detecting campaign-like abuse. Exam caveat: baselines drift when the city launches new 311 services or during disasters; freeze or segment baselines carefully and refresh after major catalog changes so detection remains calibrated. Operational check: collect prompt telemetry with privacy-minimized features, build rolling baselines for 311 AI channels, define anomaly thresholds for campaign-like bursts, route alerts to security and 311 ops, and validate detections with periodic red-team prompt campaigns.