Given a large generated test suite and a limited CI/CD time budget, this module formulates optimal test selection as a combinatorial optimization problem, solved via a quantum-classical hybrid approach — with a classical baseline always computed alongside it, both as a fallback and as the correctness check every quantum-derived result gets measured against.
The problem
Test selection is framed as Quadratic Unconstrained Binary Optimization (QUBO) — the standard form consumable by quantum optimization approaches. Each test i is a binary decision variable xi ∈ {0,1} — run it, or don't.
Where:
dᵢ— historical defect-detection rate of test i (reward for selecting high-value tests)oᵢⱼ— coverage overlap between tests i and j (penalty for redundant pairs)tᵢ— execution time of test i,B— the total time budgetλ— penalty weight balancing budget adherence against detection value
Why it needs a trust layer
The budget term above is a soft penalty, not a hard constraint — QUBO can't natively express "never exceed B," only "increasingly discourage exceeding B." This is the direct, measured cause of a real finding from testing: the quantum-derived "optimal" selection can come out more conservative than the classical hard-constraint optimum. Three factors are computed on every result before it's ever shown:
S_shot— shot consistency: the frequency of the most common result across repeated runs. Low consistency signals a noise-dominated, unreliable output.A_classical— agreement with the classical baseline: what fraction of the quantum-selected tests match the known-correct classical answer.C_hardware— hardware/backend calibration context at the time of the run.
If Trust ≤ 0.7, the system automatically falls back to the classical result — a low-trust quantum answer is discarded, not left for a human to notice.
The finding
Testing surfaced a non-obvious result: because the budget constraint is a soft penalty, the QUBO's mathematically "optimal" answer can under-use the available time budget rather than risk the penalty — producing a selection that's technically valid but leaves real defect-detection value on the table compared to the classical hard-constraint solution. This is precisely why the classical-agreement trust factor exists — it's what catches this specific failure mode automatically, rather than requiring someone to notice the shortfall by inspection.
Current scope
This runs on a local quantum circuit simulator today — it produces mathematically real QAOA (Quantum Approximate Optimization Algorithm) results against the QUBO formulation above, fully tested with 54 passing tests. Connecting to real quantum hardware is wired into the architecture but intentionally not yet enabled — an explicit next step, not implied or overstated here.
In design: FREE-AI Evidence Engine
A second research track, at the design stage and not yet built. It aims to turn the RBI’s FREE-AI framework for responsible AI in finance into continuous, tamper-evident evidence collected from running AI systems: a live AI inventory, evaluation and incident records, and auditor-ready reports.
Rules, not AI, will decide whether a control is met; AI will only summarize stored evidence with citations. It is labelled here by stage, like everything on this page, and will be published when a working demo exists.