QUASAR (Quantum Unified Arbitration System for Adaptive Recognition): A Quantum-Hybrid Architecture for Binary Supervised Classification in Payment Transactions

Instant payment systems leave little computational budget for fraud screening, and classical classifiers are least reliable exactly where transactions are most uncertain. This paper introduces QUASAR, a quantum-hybrid framework combining a LightGBM classifier with a four-qubit variational quantum circuit (VQC): a Quantum Adaptive Switch (QAS) routes only the classifier’s most uncertain transactions to the quantum module, leaving the rest to LightGBM alone. QUASAR is evaluated with a leakage-free protocol (the routing threshold is chosen on a validation set, never on the test set) on two benchmarks: a synthetic dataset and the real ULB European credit card fraud dataset. On the synthetic benchmark, QUASAR reduces false negatives relative to LightGBM alone, but at a statistically significant cost in F1 – a genuine trade-off, not a clean win. On the real benchmark, this effect disappears entirely: QUASAR is statistically indistinguishable from LightGBM alone. The standalone VQC also performs far below every classical baseline. We report the non- replication on real data, not the synthetic-benchmark result, as this paper’s central finding: evidence for caution when evaluating quantum-hybrid classifiers on synthetic data alone, rather than a case for deploying QUASAR.