Phase I Governed hybrid fraud-detection research
Can a bounded hybrid model improve the decisions that matter most?
This programme asks whether an eight-qubit residual head can improve fraud ranking inside a narrowly defined high-risk band—while controlling false positives and customer friction. Classical development evidence is complete. The hybrid comparison is designed, not yet run.
Independent research by Qonduit Limited, New Zealand. Explore the public, versioned reproducibility package. No production or quantum-advantage claim is made.
Evidence before ambition
One programme. Three distinct evidence layers.
The explorer separates demonstrated development results from accepted design choices and future qualification. That separation is the central control.
Established
Classical development evidence
Thirty completed fits across three predeclared model families, independently verified within the frozen development protocol.
See the model evidenceDesigned
Bounded hybrid comparison
A frozen eight-qubit residual reranker, matched classical heads and prospective success and falsification gates.
Inspect the architectureTo qualify
Scientific execution
The latest control-plane review requires revision. No hybrid fit, protected-test evaluation or provider run has occurred.
Follow the readiness pathM1–M2 · Data and evaluation
Chronology is part of the model.
IEEE-CIS is used as a retrospective benchmark. Equal-time groups stay together, training always precedes validation, and the protected test remains sealed.
Benchmark evidence only. Dataset behaviour does not establish operational HSBC performance or transferability.
- 1Train354,324 rowsFit preprocessing and models
- 2Model selection59,054Choose within the frozen portfolio
- 3Calibration59,054Fit calibration and thresholds
- 4Protected test118,108Sealed; no hybrid access
No future information in training
All fit decisions are made on earlier development partitions. The later segment is not folded back into training.
Protected means untouched
The protected test is reserved for a separately authorized final evaluation after configurations and seeds are frozen.
Offline is not operational
Relative transaction order supports a retrospective benchmark; it does not prove score-time feature availability or production readiness.
M3 · Demonstrated development evidence
A credible classical foundation.
XGB-003 led the bounded, predeclared portfolio on mean model-selection Average Precision. This is not a universal-best claim.
Mean model-selection AP
0.631017Best only within the bounded predeclared portfolioLater-segment signal
Every family winner declined.
Average Precision was lower on the later calibration-threshold segment for all three winners. That is a useful warning about temporal drift, not a protected-test result.
Resource observation
Bounded and measured.
Across the portfolio: 1,801 seconds of conservatively allocated active wall time, peak fit-process memory about 2.53 GiB, and maximum observed system memory of 76.3%.
Calibration boundary
Useful, but reused.
Calibration and threshold metrics use the dedicated segment that fitted those mappings. They are development artefacts, not out-of-sample operational evidence.
M4 · Frozen design
Change only the narrow band where ranking is hardest.
The proposed head cannot move a transaction across the frozen band boundary. Outside-band XGB-003 scores remain unchanged.
XGB-003 score
Base margin from the selected classical configuration.
Top 5% band
Cutoff ties enter together; the complete band is fixed per partition.
Eight-qubit head
A correction in [−1, 1], scaled so cross-band movement is impossible.
Corrected reranking
Only relative order inside the admitted band can change.
Quantum and matched classical residual heads receive the same rows, channels, base offsets, fit/monitor split and selection access.
The practical XGB-003 scores are untouched. The residual head cannot create cross-band promotion or demotion.
What “eight qubits” means here
A deliberately small, inspectable circuit.
Eight selected numeric features become eight value angles. Each qubit also receives a missingness gate, keeping “missing” distinct from “equal to the training median.”
Same-information fairness
The quantum head must beat fair alternatives—not a weak straw baseline.
Interaction logisticDegree-one and degree-two channel products
Compact tanh MLPParameter-scale matched residual heads
Random FourierBounded nonlinear residual features
No-head baselineFrozen practical XGB-003 ranking
Twelve classical residual configurations, plus the no-head practical baseline. Every learned head shares the same correction range and development information.
Prospectively frozen decision gates
Success can be falsified.
These thresholds were defined before any protected-test release. They are criteria, not achieved results.
- AP
Meaningful liftMean hybrid-minus-matched AP difference ≥ 0.010.
- 8/10
Seed consistencyAt least eight of ten paired differences are positive.
- CI
UncertaintyConservative 95% lower endpoint is above zero.
- Brier
CalibrationConservative upper endpoint for degradation is ≤ 0.002.
- Ablate
Quantum contributionA matched quantum ablation is required; hybrid lift alone is insufficient.
M5 · Implementation-readiness timeline
Controls advanced. Scientific performance did not.
The programme records corrections instead of erasing them. Passing pre-data controls do not authorize or imply model execution.
Accepted foundation
M1–M4 evidence and design
Dataset direction, chronological contract, classical development results and the bounded residual design were accepted at their stated scopes.
PASS · scopedM5 V7
Pre-data implementation controls
Fifty-seven label-free synthetic, shape and static checks passed, followed by fresh independent review with zero unresolved material findings. A launch-memory stop remained controlling.
PASS · pre-data onlyM5 V8
Control-plane review
A 151-stage deterministic control plane passed 51 bounded preflight checks. Independent review then recorded three material control gaps and confirmed that live scientific kernels are absent.
REVISE · not execution-readyNext bounded action
Add kernels, close findings, review again
No automatic resume. A hash-bound implementation of the preserved scientific kernels and a fresh independent PASS are required before any activation decision.
Future gate
Accepted classical retrospective development outcomes and an accepted hybrid design. No hybrid performance, protected-test result, prospective performance, production value, provider effect, speedup or quantum advantage.
Quantum qualification pathway
Earn the right to scale.
Qonduit Limited’s pathway starts with the smallest credible numerical instance and expands only when evidence justifies the next cost and complexity.
Exact local simulation
Resolve correctness, reachability, detectability and matched classical value under full local accounting.
Bounded simulator
Test finite-shot sensitivity with a frozen request, budget and evidence envelope.
Provider or hardware
Consider only if the numerical instance survives and end-to-end cost, latency and resource accounting are complete.
Principle Simulation cannot establish hardware advantage. A provider run cannot rescue a weak or non-defensible numerical hypothesis.
Phase II decision path
A useful outcome in either direction.
Phase II should begin only after the implementation reaches a clean, independently reviewed activation point.
Does the hybrid hypothesis survive the frozen tests?
Evaluate practical lift, seed consistency, uncertainty, calibration, ablation and total resource cost together.
Bounded hybrid pilot
Advance the smallest validated configuration. Preserve the band, monitor false-positive impact, and maintain a classical fallback.
Classical-only exit
Retain XGB-003 as the practical benchmark, document the falsified hybrid hypothesis, and redirect effort to data quality, calibration and drift.
Read the boundary
What this explorer does—and does not—show.
What is demonstrated
A chronological retrospective evaluation design and development-only classical results within a bounded predeclared portfolio.
What is designed
An eight-qubit residual reranker, twelve matched classical residual comparators, a no-head baseline, and frozen success/falsification gates.
What is not demonstrated
Hybrid gain, quantum contribution, protected-test performance, prospective or HSBC operational performance, savings, deployment, provider effect, speedup or hardware advantage.
Why the boundary matters
Governed research becomes decision-useful by making uncertainty, revisions and exit criteria visible—not by turning a design into a result.
Governed hybrid research
Start with a falsifiable question.
Qonduit Limited helps teams frame, test and account for hybrid and quantum opportunities without skipping the classical benchmark or the evidence gates.