Quantum decision systems: from experiment evidence to investment gates.

IBM's public quantum technology roadmap describes Heron processors, Quantum System Two direction, and a path toward near-term advantage. That does not remove the engineering problem: quantum decision systems need evidence that can be checked, repeated, and compared against classical baselines.

QANTIS and qmesh should therefore be presented as evidence systems. The value is not only a circuit or benchmark, but the full chain of assumptions, execution metadata, mitigation choices, and decision quality.

Classical planning, simulation, and policy logic remain responsible for reliability. Quantum execution is inserted where it can be measured against an explicit decision problem.

The evidence record should therefore preserve the classical formulation, backend and calibration context, circuit or workload version, mitigation choices, resource estimates, output distribution, and fallback decision together. Reviewers can then ask whether the quantum step changed decision quality or cost under the tested conditions, rather than treating provider access or circuit completion as proof of advantage.

Most evidence failures are not bad experiments. They are missing links between good artifacts. A run executed without captured backend metadata cannot be repeated under comparable conditions, so its result decays into an anecdote. A mitigation strategy that was applied but never versioned makes two output distributions incomparable even when everything else matches.

Baseline drift is the quieter failure. The classical baseline improves or degrades over time while archived quantum results stay fixed, and every comparison made against the stale baseline is silently invalid. The fix is procedural, not technical: version the baseline with the same discipline as the quantum record, and re-run the comparison when either side changes.

The last failure is silent fallback. A system that reverts to its classical path without recording the event produces decision logs that attribute outcomes to quantum components that never ran. Fallback must be an observable, logged state, visible in telemetry and in the operator interface.

The first instrumentation point is not the quantum program. It is the boundary where a quantum output enters a decision. If that boundary is not logged, everything upstream is unattributable no matter how carefully the physics was recorded. Start there and work backwards.

Working backwards, the chain assembles in the order this article describes: problem definition, simulator pass, backend metadata, mitigation strategy, output distribution, decision impact. Each link should be a required field in the run record, not an optional annotation. A run missing any link should be marked incomplete by default.

The Neura Parse stack maps onto this order. qmesh provides the substrate layer, with ed25519-signed manifests and offline-verifiable hash chains that pin what was executed. QANTIS structures the decision side, with a Decision Engine organised as Infer, Risk, Optimise, Verify, giving each evidence artifact a defined stage to attach to. QFlow Studio is where the assembled chain becomes a reviewable record: assumptions, baselines, resource estimates, and provider context in one place.

A benchmark number without its evidence chain is a marketing artifact. The evaluation question is not what score a vendor reports but whether the run behind it can be checked, repeated, and compared against a stated classical baseline. Vendors who can produce the chain will produce it. Vendors who cannot will offer a slide.

The same standard applies in the other direction. The public QANTIS repository is the Community Edition under MIT, a didactic surface, while public benchmark numbers and hardware campaign results are bounded by the linked arXiv records; additional private artefacts are discussed through formal engagements. That separation is not modesty. It is the evidence discipline applied to our own claims, and it is the posture a program should require of every supplier.

IBM's public roadmap points toward more capable hardware and near-term advantage. Each hardware generation will invalidate old benchmark comparisons, because those numbers were measured on machines that no longer represent the frontier. What survives generational turnover is the evidence discipline itself: the problem definitions, baselines, and comparison methodology carry forward even when the results do not.

This is the practical argument for hybrid architecture as a standing posture rather than a transition phase. Classical planning, simulation, and policy logic keep the reliability responsibility, so quantum components can be swapped as hardware improves without re-deriving trust from scratch. A program with intact evidence chains can re-run its decision problems on new hardware and read the delta directly.

The organizations that keep a classical baseline close to every quantum result will be the ones able to say, precisely and with records, when quantum insertion started to pay. Everyone else will be arguing from vendor slides. That is the position worth building toward now, while runs are still cheap enough to instrument fully.