CAST-DRO: Cognition-aware service triage with configurable risk adjustment for resource-constrained AI information services

Abstract

AI information service platforms must allocate heterogeneous requests across models, memory, clarification, and human review under joint resource constraints. We develop CAST, a cognition-aware five-action service-triage framework; CAST-ERM learns group-level empirical allocations, and CAST-DRO adds a configurable mean-plus-variability adjustment. Using sequence-grouped train/evaluation splits for 520 LifeSim-derived evaluation requests across 20 repeated splits, CAST-DRO improves over rule-based CAST in quality (0.679 to 0.743), failure rate (0.617 to 0.350), and final satisfaction (46.35 to 56.51), while reducing normalized high-value resource cost (0.087 to 0.081). CAST-ERM reaches quality 0.738, confirming that cognition-group empirical allocation explains most of the average gain. The risk adjustment changes 12 of 104 eligible group-seed action selections and sharply reduces selected-action loss variability in the medium-risk group. A LinUCB comparison warm-started with full-information simulator outcomes reaches quality 0.744 with high-value cost 0.092. A strict budget audit finds zero human, memory, or total-spend violations for all four hard-budget policies across all 20 splits. Resource-price and capacity tests retain a positive quality advantage over CAST, although high-value resource use depends on the bottleneck. In a two-layer stress test, separate mandatory safety-review capacity of 30% reduces latent-reference high-risk miss from 0.788 to 0.230, with 0.035 overflow and 0.398 total escalation. Real-model evaluation on 100 requests and anonymized 1–5 ratings of 50 request–response pairs provide supporting, not production-level, validation. CAST-DRO is best interpreted as an auditable middle-layer allocator after mandatory safety filtering.