SynQubi: An Agentic Framework for Trustworthy Quantum–Classical Financial Decision-Making with Generative Tail-Risk Scenarios

Hybrid quantum-classical machine learning is increasingly tested in security-sensitive financial settings, yet evaluations rarely make it possible to attribute observed behavior to the quantum components themselves. We present SynQubi, a modular agentic framework for risk-aware financial decision-making in which each quantum component occupies an isolated, auditable role and is benchmarked against strong classical references. We instantiate two agents end-to-end: a Scenario Generator based on a Quantum Circuit Born Machine (QCBM) that learns the joint copula of credit-spread movements and samples coherent tail scenarios, evaluated against empirical, Gaussian, and Student-$t$ copula baselines; and an Optimization Agent based on the Quantum Approximate Optimization Algorithm (QAOA), whose portfolio proposals are certified against exact classical solvers. The two agents are connected through a scenario-conditioned risk evaluator: generated tail scenarios reshape the risk term of the downstream allocation problem, allowing us to measure whether, and when, the choice of quantum scenario model changes the final decision. Across seeds we report distribution-fit metrics, joint-tail probabilities, optimality gaps, and feasibility rates: on non-degenerate instances the QAOA component matches classical references on solution quality and is markedly worse on constraint satisfaction, producing budget-feasible allocations in roughly half of samples against the null model's all. Turned on our own pipeline, the protocol overturned five results that ordinary inspection had passed, and their directions differ: a units error erased a real cross-source effect; an inconsistent pricing mechanism manufactured a spurious one; a degenerate benchmark whose certified optimum was a single asset reversed the headline optimizer verdict; a truncated export left the scenario generator unable to emit the tail states it exists to supply; and a joint-tail match credited to the circuit decomposes into an empirical mixture weight. Artifacts with no common sign cannot be corrected by discounting favorable results, only by a mechanical reference. We establish no quantum advantage, which is the condition under which these diagnostics are informative: in high-stakes financial deployment the primary question is not whether quantum components outperform classical ones, but whether the surrounding system makes it possible to tell.