Proceedings of the Annual Meeting of the Cognitive Science Society

Detecting rare, meaningful signals in non-stationary, heavy-tailed environments challenges biological agents and artificial monitors alike. In network security, analysts must identify sparse super-spreaders within vast benign traffic; as baselines drift, fixed thresholds either overload attention with false alarms or miss emerging threats. We present CoSALT, a resource-rational anomaly detector that frames criterion setting as a Minimum Description Length (MDL) problem. CoSALT adaptively partitions observations into a compressible bulk and a salient tail, while an explicit cognitive selection cost penalizes attending to too many candidates. It further uses spline-based quantile regression to learn context-dependent sparsity expectations. On real backbone-traffic benchmarks, CoSALT improves F1 relative to strong baselines. In a behavioral study, its selection boundaries closely match human visual judgments, especially under explicit budget constraints. These findings are consistent with a resource-rational account of adaptive detection in dynamic, resource-constrained environments, while leaving room for alternative cognitive mechanisms.