When does “hybrid quantum-classical” hide the classical doing the work?

"Hybrid quantum-classical" is the hardest-working phrase in quantum computing right now, and it is doing too many jobs at once. It names at least three different architectures, and it routinely gets read as "the quantum part did the work" when nothing in the label says that. Rosetta Quantum treats the word as an audit prompt, not an answer: this post lays out the taxonomy, the numbers from the decomposition audits that exist as of September 10, 2026, and the four questions we ask before a "hybrid win" earns a row in our ledger.

What does "hybrid quantum-classical" actually mean?

At least three different things — and the difference decides who did the work. In today's usage the label covers: (1) the variational loop, where "a classical optimizer [is used] to train a parametrized quantum circuit" (Cerezo et al., Nature Reviews Physics 3, 625–644, 2021) — VQE and QAOA are the canonical cases; (2) the partition pipeline, where a classical workflow decomposes the problem and hands selected fragments to a QPU — D-Wave's Leap hybrid solvers are the commercial reference; and (3) classical pre/post-processing around quantum samples, where the quantum device produces samples and a classical solver does the heavy numerical lifting — IBM's sample-based quantum diagonalization (SQD) is the documented example. QuEra's public glossary (accessed September 10, 2026) frames the paradigm as "a classical computer and a quantum computer work[ing] in tandem," with the classical side managing optimization — a fair definition that still leaves the work split unstated.

None of the three is a trick. All three are legitimate engineering, and vendors document them openly. The honesty problem is downstream: one word covers all three, and a result reported under that word tells you nothing about which side produced it.

What did the one decomposition audit actually measure?

That in the audited runs the QPU was active for 0.034 seconds of a 5-second nominal budget — about 0.68% of the budget and ≈0.7% of measured wall-clock, with the remaining ≈99% spent in classical decomposition and reassembly (Lozano, "Where the Quantum Lives in D-Wave Hybrid Portfolio Optimization: An Operational Decomposition Audit," arXiv:2605.17623, June 2026). The audit ran cardinality-constrained portfolio problems from N=10 to N=640 assets through D-Wave's hybrid CQM and BQM interfaces against Gurobi and simulated annealing.

Two more numbers from the same audit matter as much as the headline share. First: the hybrid CQM matched Gurobi's proven optimum on all 54 head-to-head instances at N≤120 — solution quality was not the failure. Second: a purely classical TabuSampler reached CQM-equivalent objectives within a mean absolute gap of 0.001 (max 0.008). The author's conclusion is the methodological point: the reported hybrid wins reflect "constraint-native classical pipelines" rather than quantum-sampling advantages at the tested scales. The pipeline is good. The attribution is the problem.

To be fair to the vendor: D-Wave's own documentation describes the Leap hybrid solvers as classical-quantum workflows and has never claimed the QPU does most of the wall-clock work; and matching a proven optimum at N≤120 is a respectable result for any heuristic. The audit's finding is about where the credit goes when the word "quantum" appears in the result — not about hidden misconduct. That is exactly why the decomposition, not the label, should be the reporting unit.

Is the variational loop hiding anything?

Less than the other two classes — its split is declared by construction — but it has its own version of the problem. A variational algorithm openly consists of a classical optimizer plus quantum circuit evaluations; nobody hides the loop (Cerezo et al., 2021). The audit question there is whether the quantum evaluations are contributing search power or just noise for the classical optimizer to fight through: the same review names trainability — barren plateaus included — as an open challenge, and QuEra's glossary lists it as a standing limitation of the paradigm.

The freshest measured data point is not flattering. A July 2026 benchmark of hybrid QAOA-based methods on power-grid unit commitment (AC-OPF-UC, 5 to 13 generators) found that the qubit-efficient hybrid "does not outperform uniform sampling" — the pipeline produced valid answers, but replacing the quantum sampler with uniform random draws did just as well at the tested sizes; the authors state that testing potential advantage would require 25+ generators, beyond their study (Gaidai & Mukherjee, arXiv:2607.15543, July 17, 2026). A loop that runs is not a loop that helps. Rosetta has published the mechanics of both loop algorithms before — see our QAOA and VQE explainers — and the verdict standard is the same here: the loop earns credit when an ablation shows the quantum samples beat a dumb sampler at equal budget.

When does classical post-processing carry the result?

When the pipeline's own documentation says so — and for SQD it does. IBM's documentation describes the workflow plainly: the quantum computer prepares states and produces samples; then a classical loop performs configuration recovery and "the Hamiltonian is projected and diagonalized in the subspace spanned by those samples," repeating to convergence; and "the computational cost of SQD is dominated by the eigenstate solver calls" — which run on classical, embarrassingly parallel infrastructure (IBM Quantum documentation, accessed September 10, 2026). That is not a leak; it is the published design. The open question is what the quantum samples add over classically generated configurations — and how the subspace scales.

On the second question there is now a measured negative: a critical assessment on Heisenberg and Hubbard models found that the number of configurations needed to reproduce the ground-state energy within fixed accuracy "grows exponentially with system size," even under optimal inclusion of configurations in order of decreasing probability (Gaberle & Jattana, arXiv:2605.02494, May 2026). That result is model-specific — it will not transfer automatically to every chemistry target — but it is exactly the kind of number the word "hybrid" was hiding: if the classical diagonalizer needs exponentially many configurations anyway, the quantum sampler is not the engine of the result.

How does Rosetta audit a "hybrid" claim?

Four questions, in order, before any "hybrid win" enters the Rosetta Quantum ledger — each one already has a precedent in the sources above:

1. What is the wall-clock split? Ask for QPU seconds over total seconds, the way the Lozano audit reported it (0.034 s / ≈4.7 s measured). QOBLIB's submission rules — which we covered in our benchmark explainer — already require total wall-clock declaration; a hybrid claim should itemize it.

2. What happens under ablation? Replace the quantum samples with a cheap classical source (uniform random, tabu, greedy) inside the same pipeline. The grid benchmark's uniform-sampling baseline and the audit's TabuSampler swap are the template. If quality holds, the pipeline — not the QPU — is the story.

3. Is the classical baseline the same pipeline minus the QPU? Comparing a polished hybrid pipeline against a weak standalone classical solver is the oldest trick in benchmarking — we wrote the weak-baseline anatomy in July 2026. The fair control is the vendor's own classical machinery with the quantum call stubbed out.

4. Who published the decomposition? If the split comes from the vendor's own docs (as with IBM's SQD cost statement), credit the disclosure. If no split is published anywhere, the claim is unaudited — Rosetta files it as "pipeline win, attribution pending," never as quantum advantage.

What we know / what we don't know

What we know (measured, sourced above): "hybrid" currently names at least three architectures; in the one third-party wall-clock decomposition audit published this year, the QPU accounted for 0.034 s of a 5 s budget (≈0.7% of measured run time) and a classical tabu search matched the hybrid's objectives within 0.008; a hybrid QAOA grid pipeline did not beat uniform random sampling at 5–13 generators; and IBM's own documentation states SQD's computational cost is dominated by the classical eigenstate solver.

What we don't know: whether these splits generalize — the D-Wave audit covers one problem family on one vendor's interfaces, and its author notes the QPU could contribute differently at other scales or problem structures; whether class-1 loops help above the tested sizes (the grid authors themselves say 25+ generators is the open test); whether SQD's exponential-configuration result on Heisenberg/Hubbard models transfers to the chemistry systems where SQD is actually deployed; and — our own limit — Rosetta Quantum has not yet re-run any of these decompositions in-house. Everything here is a reading of the public record with sources attached, not an independent re-measurement. When we run our own hybrid ablation, it will be published with its seal, whichever way it lands.

Rosetta Quantum publishes verdicts with raw, reproducible data. This is educational content — an audit method, not a product claim and not procurement advice. Quantum, independently measured.