Verified Evidence Registry & Claim Boundaries

Q-Psi demonstrated compiler-enabled physical quantum query advantage on structured candidate-state search problems on IBM Heron hardware under an opaque black-box verifier model (N=4: 2.43x, N=8: 1.73x, N=16: 1.19x; 9/9 individual cases; 3/3 tested sizes).

Q-Psi demonstrated statistically significant physical quantum query complexity scaling advantage on the Dynamic Bernstein-Vazirani problem across N=4..16 qubits on IBM Heron hardware (slope 0.153 vs classical 0.696, t = -30.65, p < 1e-6).

Compiler-driven liveness and mid-circuit reuse reduced physical register width by a median 78.1% (up to 87.5% on 16 logical qubits mapped to 2 physical slots) across the frozen streaming benchmark while all six tested cases remained within the predeclared observable-quality criterion.

Q-Psi demonstrated a physical-QPU-assisted coding-data pipeline in which QPU measurement samples causally selected repository mutation states that were converted by executable software verifiers into 93 unique verified repair records (127.58s runtime, 332,768 shots, 100% QPU provenance).

On standard MQT benchmarks on IBM Heron, Q-Psi compiled all 18 frozen benchmark cases. Physical competitiveness was workload-dependent: Qiskit held a small aggregate fidelity advantage (-0.0067 delta), while the apparent aggregate advantage over TKET (+0.0398 delta) was driven by one routing-sensitive workload (vqe_su2_n8).

Dynamic routing compressed 2Q depth from 73 to 2 on the longest tested path (13 hops), but hardware reset and measurement noise prevented physical Bell-fidelity crossover on Heron r2, confirming pre-QPU compiler predictions.

QAOA-based compiler optimization on IBM Marrakesh achieved 4/4 exact optimum hits on small compiler graphs (N<=10), matching classical branch-and-bound, but degraded on N=18..25 due to hardware noise.

Initial proof-of-concept demonstration of physical QPU selector on M=1024 space yielding 14 verified unique code repair records (4,096 shots, 4.70s runtime).

Against a strong baseline of Qiskit O3 best-of-8 seeds, Q-Psi calibration-aware layout achieved 1 win, 2 losses, 3 inconclusive (mean delta -0.0106), demonstrating that Qiskit multi-seed search remained superior or competitive on Heron r2.

Fractional gate lowering yielded structural pulse duration reductions and 2/6 physical wins, with 4/6 cases inconclusive on IBM Marrakesh, supporting mixed physical signal.

Selective Pauli twirling twirled only 41.1% of CZ layers on IBM Directed Execution beta mode (58.9% reduction in twirling overhead); achieved 1 win vs unmitigated RAW baseline; comparison vs full twirling remained inconclusive.

Physical execution on IBM Marrakesh recovered secrets on 6/12 instances (N=8..28) with polynomial query count reduction relative to classical bound, but noise-induced degradation on larger N prevented establishing asymptotic quantum speedup.

Exploratory hardware calibration run on static BV and mantra state encoding on IBM Marrakesh demonstrated distinct measured output distributions.

Across six canonical workloads and three IBM Heron r2 backends, Q-Psi successfully retargeted all 18 workload-backend combinations. Its pre-QPU selection heuristic placed the selected backend in the physically best statistical tier for 5 of 6 workloads, with one selection miss.

General quantum advantage for arbitrary classical software or compilation tasks is NOT established by Q-Psi research and is explicitly disclaimed.

General compiler runtime or optimization superiority over industrial compiler toolchains (Qiskit, TKET) is NOT established by Q-Psi research and is explicitly disclaimed.