A Two-Agent Triage Framework for Flagging Uncertain Predictions in Child Welfare Case Prioritization

Predictive models are increasingly proposed to help resource-constrained child welfare agencies prioritize cases, but a single confident-but-wrong prediction can be more harmful than an admission of uncertainty. We propose a lightweight two-agent triage framework in which an independent verifier agent, trained on a disjoint feature view, cross-checks a primary predictor's output; disagreement or low joint confidence automatically routes the case to human review rather than to an automated decision. On a simulated case-outcome dataset structured after the U.S. AFCARS foster care schema, our two-agent system with class-balanced training raises F1 on the subset of cases it remains confident about from 0.471 to 0.629, and reduces unflagged false negatives (cases confidently predicted stable that were actually prolonged/unstable, and therefore never routed to a human reviewer) from 82 to 49 out of 750 test cases, while flagging 52.3% of cases for human review. We position this as a proof-of-concept architecture, not a validated clinical or casework tool, and discuss the real-data access, fairness, and deployment challenges that must be resolved before any such system could responsibly touch real cases.