THE QUANTUM INTELLIGENCER

The most instructive event in the season's record of post-quantum preparation is not a migration deadline but a defect that no deadline would have exposed. A report posted to the arXiv, and not yet passed through peer review, recounts the design of an accelerator for the ML-KEM-768 and ML-DSA-65 algorithms, driven from register-transfer to PCIe bring-up on a Xilinx Kintex-7 XC7K160T by an agentic language model across two hundred and thirty-two logged experiments. The device cleared its full known-answer test regression while carrying a norm check that outran the block-RAM latency, leaving each signing candidate's final coefficients unverified; the fault appeared at the fifth rejection-loop iteration. The authors locate the blind spot in the instrument, not the author: the standard acceptance gate samples fixed values, whereas the signing path under scrutiny varies with the message, so no fixed vector could reach the failing depth. Their corrective gate, a byte-exact golden-reference oracle paired with randomized adversarial soak, closed the gap over three hundred and one thousand three hundred and forty-three data-dependent signings with no escapes. That the artificial author succeeded in but 71.6 per cent of its experiments does not establish that an unreliable author may be trusted; it establishes, if anything, that the test that accepted its work was the weaker link. The report is a preprint, and its figures are the authors' own; prudence will treat them as the claims of a not yet independent study. The report's tables carry further figures that bear transcribing. The accelerator shipped at 98.5 per cent slice occupancy on its Kintex-7 XC7K160T. Success rates followed the gradient the authors describe: documentation and research tasks succeeded in 77 to 85 per cent of runs, while synthesis and bring-up succeeded in only 50 to 53 per cent, a gap they attribute to observability, since failure concentrates where corrective signals are physical-side only. The deployed baseline survived 779,945 checks of the randomised adversarial soak with zero failures, remaining byte-exact across all six FIPS operations. These numbers, like the rest of the report, are claims from a preprint awaiting independent review. Let no reader mistake the extent of this report. Its title calls it a case study, and in scope it is precisely that: one accelerator, one FPGA family, two cryptographic algorithms, one authoring agent. The 232 experiments are 232 runs of a single design effort, not a survey of the art. The 71.6 per cent figure is a descriptor of that one effort, and attaching it to the broader question of artificial intelligence authorship would be an error the document does not commit. Neither does its success rate, nor its single escape and the gate that closed it, license an inference about other chips, other algorithms, or other agents. What the report establishes is that a particular norm check in a particular design outran a particular memory latency, and that a particular test then missed it. If it speaks at all beyond that particularity, it does so by example: of the kind of error that acceptance fixtures do not reach. That is a narrow authority, but it is authority nonetheless, and a prudent engineer will keep it in its lane. —Ada H. Pemberley Dispatch from The Prepared E0