Fraud detection is a hard place to make an easy claim.
A model can report excellent accuracy while detecting no fraud. A randomly split benchmark can leak the future into the past. A threshold that looks good in a paper can generate an unusable alert queue. And a “quantum” component can consume most of its apparent value in feature loading, sampling or post-processing.
The HSBC challenge research in the Haivian Research Suite started from those risks. It used public benchmark evidence: no HSBC customer data, production deployment, policy, validation or endorsement is claimed. Its first deliverable was not a quantum model. It was an evaluation contract designed to make a later experiment falsifiable.
Why accuracy is not the decision
Consider a dataset in which fewer than one transaction in 100 is fraudulent. A classifier that labels every transaction legitimate would report accuracy above 99% while detecting no fraud at all. This is a mathematical illustration of class imbalance, not a measured result from the HSBC work or a claimed statistic for a particular dataset.
The illustration makes the governance point immediately: an attractive headline metric can hide complete operational failure.
The public data and evaluation record uses an integrity-first hierarchy. Provenance, split design, leakage controls and an untouched test set come before any score. Average precision—how well a model concentrates real frauds near the top of its ranking—is more informative for a rare positive class than accuracy alone. Confusion counts, calibration—whether predicted risk matches observed frequency—alert rate and threshold consequences then show what the score means at a decision point.
This is not an HSBC production policy. It is the research programme’s proposed contract for a public benchmark challenge.
The comparator has to be real
The programme treated “at least one classical baseline” as a floor, not a persuasive comparison. The public classical-development record documents the bounded comparator state. A credible experiment needs strong tuned classical models, the same information boundary and the same temporal partitions as any quantum or hybrid candidate.
For a public fraud study, the research plan required preprocessing and feature selection to be fitted only on the historical side of each split. It also separated probability estimation from threshold selection: a model score is not an operational decision until an authorized cost or capacity rule chooses a threshold.
The purpose of these controls is not to make quantum methods fail. It is to make a positive result worth believing.
A public design that had not run
The public hybrid-design record describes an accepted design for a bounded hybrid re-ranking study. In ordinary terms, the proposed component would reconsider only a narrow band of close cases after a frozen classical comparator; it would not replace the fraud model. The public source-evidence map supplies content-addressed identities and the scientific claim ceiling; it is an index and hash register, not an archive of the named source records.
The same evidence map explicitly limits the claim: the design was accepted, but scientific hybrid execution had not been completed. That is a design milestone, not a hybrid performance result.
Keeping those categories separate prevents a proposed experiment from being reported as an executed one.
What a publishable result would still need
A future claim that a bounded quantum or hybrid component improved fraud ranking would need, at minimum:
- Evidence: an exact workload identity, chronological leakage-resistant partitions and a protected test boundary.
- Comparison: a strong same-information classical residual head, a no-head ablation, one predeclared primary metric and uncertainty analysis.
- Decision use: calibration and operating-point evidence, including alert volume and threshold consequences—not a ranking score alone.
- Verification: complete preprocessing, loading, simulation, reconstruction and checking costs, plus reproduction under the protected boundary and a clear statement that a public benchmark is not HSBC production validation.
The current record establishes a research question and a governed route to test it. It does not establish quantum advantage, improved fraud detection, production readiness or performance on HSBC customer data.
Evidence boundary
The readiness view provides the public boundary for what has and has not run; the source-evidence map above controls the permitted claim ceiling. Both were recorded for this candidate as at 2026-09-28.