Loss-Responsive Dual-Memory Graph Models for Fraud Drift: A Controlled Temporal Stress Test

Adaptive graph fraud detectors must decide both when to update and how strongly to trust an updated model. This study isolates those decisions through a fully reproducible synthetic temporal-transaction benchmark. Following the ATGNN-SP design, we implement directed two-hop simplified graph propagation, a frozen stable model, an incrementally updated plastic model, a validation-calibrated loss gate, and anchor replay. We extend the fixed stable–plastic mixture with a delayed-label exponential expert-weighting rule. Across five seeds and 15 future snapshots, the adaptive mixture attains average precision 0.4170 and F1 0.4021, versus 0.3469 and 0.3189 for a fixed mixture; however, it updates on all 15 snapshots, compared with 7.8 for the fixed mixture. Plastic-only adaptation is competitive (AP 0.4163; F1 0.3965), showing that retaining a frozen predictor in the final score need not always improve outcomes. All results are computed from a synthetic generator with deliberately changing label and connectivity mechanisms; they should be interpreted as a controlled stress test, not empirical validation on banking or cryptocurrency transactions.