Quantum research, measured end to end.

A quantum method has not established a useful result until it is compared with the strongest classical method we can field on the same workload: exact or convex solvers where they apply, MILP and CP-SAT, greedy and local search, tabu, and simulated annealing. Apple Silicon GPU acceleration through PyTorch MPS can support local classical experiments. Provider access implies no partnership, endorsement, exclusivity, or measured improvement.

Benchmarks report the whole pipeline, not a flattering slice: feasibility, objective value, solution diversity and stability, total wall time, queue, embedding, and transpilation overheads, processor time, monetary cost, and — where the workload feeds a decision — downstream regret. A processor-time improvement alone does not establish an end-to-end win; quality, feasibility, stability, total wall time, cost, and downstream decision impact must be reported together.

Candidate workloads are discrete and decision-shaped: representative branch selection, robust action selection, assignment, and cardinality-constrained hedge selection. Continuous work — scenario generation, football trajectory generation — remains classical.

We make no quantum-advantage claim. A method earns a product path only if it improves a decision-relevant measure under an end-to-end comparison; otherwise the strongest classical method remains the production choice. Benchmark artifacts are in preparation and will be published only with their configurations, seeds, costs, and limitations.