Building a Governed Foundation for Hybrid Fraud-Detection Research

Three evidence layers, kept separate

The programme distinguishes what has been demonstrated from what has been designed and what still requires scientific qualification. Classical development evidence has been produced and verified. The bounded hybrid design has been accepted and frozen. No hybrid fit or protected-test evaluation has taken place.

This separation matters because an experimental design is not a performance result. It lets the next phase ask a precise question without presenting an intended comparison as an achieved gain.

A chronological benchmark foundation

The accepted Phase I evidence uses the retrospective IEEE-CIS benchmark of 590,540 labelled transactions: 20,663 fraud and 569,877 non-fraud. The programme froze chronological partitions of 354,324 training rows, 59,054 model-selection rows, 59,054 calibration rows and 118,108 protected-test rows.

The protected test remained outside development. That boundary keeps model construction and selection separate from any later confirmation, while complete timestamp groups remain together across the chronological split.

These are benchmark and development results. Dataset behaviour does not establish HSBC operational performance or transferability. Operational feature availability, label timing, customer impact and production conditions require separate evidence.

What the classical development evidence shows

Qonduit completed 30 fits across three predeclared model families. Within that bounded portfolio and development budget, XGB-003 led with mean model-selection Average Precision of 0.631017 and mean ROC-AUC of 0.928914.

Those values describe the frozen development protocol. They are not protected-test results, an exhaustive search or evidence of performance in HSBC operations. The classical model provides a stable base and a practical fallback for the proposed comparison.

A bounded eight-qubit residual head

The accepted design freezes XGB-003 as the base model and permits a residual head to rerank only the highest-scoring 5% of a retrospective batch, including cutoff ties. The correction is mathematically bounded so that no transaction crosses the band boundary.

The proposed eight-qubit head receives the same information as twelve matched classical residual alternatives. A no-head baseline tests whether adding any residual component is useful. Improvement over the base alone could not identify a quantum contribution; the matched controls are essential to that question.

The design treats ranking, false positives, calibration and customer friction as connected parts of the evaluation. A score improvement cannot silently override the frozen calibration, uncertainty and operational-effect gates.

Phase II is a falsifiable decision path

Phase II must first qualify the numerical instance and the information parity of every branch. Only a qualified experiment can proceed to the frozen comparative gates. Any failed validity, parity, resource or mandatory performance gate ends the hybrid route or makes the result inconclusive.

The useful outcome is therefore a defensible decision: continue investigating a promising bounded hybrid approach, retain a classical-only solution, or report an inconclusive result. A classical-only or inconclusive outcome remains valuable because it narrows the decision space without overstating what quantum methods contributed.

What has not been demonstrated

No hybrid fit, protected-test evaluation, provider or hardware run has occurred. The current record establishes no hybrid gain, quantum contribution, HSBC production result, transferability, production value, financial saving, speedup, deployment readiness or quantum advantage.

Those limits are part of the public result. They preserve the difference between a governed experimental foundation and a claim that the future experiment has already answered.

Explore the evidence

The HSBC Decision Explorer presents the demonstrated classical evidence, frozen hybrid architecture and future qualification path in an inspectable public form. The reproducibility package provides the public reference aggregates and checks without including the underlying competition data.