We introduce a role‑aware Decider‑Supervisor (DS) framework with blockchain auditability to examine when a second AI component can improve primary network‑fraud decisions instead of adding operational overhead. The framework evaluates four directional configurations: a centralized machine‑learning model, a federated meta‑model trained by Federated Averaging (FedAvg), a Base large language model (LLM), and a Quantized Low‑Rank Adaptation (QLoRA) LLM variant.
Primary‑only decisions are compared with supervised decisions across metrics such as non‑hard fraud performance, intervention burden, conditional calibration, traffic‑mix sensitivity, review‑capacity sensitivity, dependability tests, and blockchain lifecycle controls. A deterministic hard gate resolves 89.994% of fraudulent requests, leaving the non‑hard population as the main AI decision setting. Conditional validation calibration does not yield a consistently transferable supervisory advantage in deployment replay. DS‑3 (QLoRA) is the least disruptive supervised configuration, yet it underperforms its preceding FedAvg stage in F1 score and total errors. Across 36 reweighted traffic mixtures, supervision reduces total errors only for DS‑4 (Base) in two extreme high‑fraud scenarios.
Blockchain experiments confirm digest verification, tamper detection, authorisation, single‑use review resolution, and post‑finalisation integrity, while exposing a pre‑finalisation single‑write limitation. The results indicate that the value of AI supervision depends on role assignment, calibration, escalation policy, traffic composition, and lifecycle controls rather than merely the presence of a second model.
Review