THE QUANTUM INTELLIGENCER

A preprint posted to the arXiv service addresses a defect in the manner quantum transpilers are judged. The customary oracle compares compiled output against a reference: a statevector, a sampled distribution, or a unitary compared modulo global phase. A companion empirical study reports that roughly 28 per cent of merged Qiskit transpiler bug-fixes, with a 95 per cent Wilson confidence interval of 19 to 40 per cent, repair a fault that corrupts layout metadata, global phase, or run-to-run reproducibility while the output remains correct; such faults are invisible to a black-box output-equivalence oracle by construction. Against them the authors set a fault-class-matched oracle family: a layout and permutation contract checker for the metadata channel, a contract-level metamorphic relation, a global-phase tracker, and a determinism runner. Verification from source code, across nine real, merged Qiskit transpiler regressions, three per channel, found the output-equivalence oracle blind throughout and the matched mechanism firing on every case. A sweep of 675 configurations of the contract and metadata invariant produced no false positive; synthetic mutant families gave 1.00 sensitivity and specificity across 36 mutants apiece for the contract and phase channels, and 1.00 sensitivity, with a confidence interval from 0.44 to 1.00, on the three circuits for which the reproducibility mutation could be constructed. The contract checker costs two to six orders of magnitude less than a plain output check; the global-phase tracker is comparably cheap within its tier; only the metamorphic relation carries a bounded cost. Ported natively to pytket/tket, the global-phase mechanism transferred cleanly, with phases recovered to double-precision accuracy, evidence that the result is not a Qiskit artifact. This is a preprint, not a settled finding. It does not establish that output-equivalence oracles are without use, nor that the matched family names every fault class; it establishes a measurable blind spot and a mechanism that covers it. The scope of the result should be kept in view. This is a paper about one component of the toolchain, the transpiler, and within that about the regression oracle, not a certificate for quantum compilation at large. Nine real regressions were verified from source; that number is what the material allowed, not a claim that the fault universe contains nine members. The confidence interval on the reproducibility channel, 0.44 to 1.00, is wide because only three circuits admitted the constructor, and the authors do not pretend otherwise. The port to pytket/tket widens the claim modestly: two transpilers are a beginning, not an atlas. A narrow result, stated as such, and none the worse for it. —Ada H. Pemberley Dispatch from The Prepared E0