THE QUANTUM INTELLIGENCER

TEL AVIV, 21 AUGUST — TEL AVIV, 21 AUGUST — Quantum X Labs, the Tel Aviv concern, reports that its AI-driven decoder has demonstrated improved performance against matching-family benchmarks on a public surface-code configuration drawn from Google's real-hardware experiment, while being trained on synthetic samples alone. Prof. Nir Sharon, the company's chief quantum scientist, called the result a validation point for its roadmap toward trusted quantum error correction, while cautioning that this is one benchmark configuration and that replication across device centres remains the next objective. The announcement, made this morning, indicates that the migration of error correction from simulation to operational hardware is proceeding, but the distance between synthetic training and real-world behaviour remains measurable. The announcement, filed from Tel Aviv this morning, reads less like a benchmark report than a declaration of intent. It records a milestone: a public surface-code configuration, drawn from Google's real hardware, was made to yield improved results by a decoder trained on synthetic samples alone. But the margin is not stated. No percentage of improvement, no error-rate reduction, no comparison of logarithms; only the assertion that the matching-family results, including Google's own correlated matching and PyMatching, were exceeded. The omission is worth noting at length because the absence of a figure is itself a datum: the company is content to advertise direction without magnitude. The drama of the times is elsewhere. Craig Gidney, writing on 24 December, chronicled the half-life of logical qubits under error correction: one hundred microseconds in 2014 with nine qubits; three milliseconds in 2021 with twenty-one; three hundred milliseconds in 2023 with fifty-one; two hours in 2024 with fifty-nine. A leap from three hundred milliseconds to two hours, in a single year, is the sort of improvement that makes fixed benchmarks quarrels. It is also the sort of improvement that invites exaggeration in adjacent claims. IBM and the University of Chicago, in July, reported a demonstration on seventy logical qubits: 2,415 logical two-qubit operations, 468 logical T gates, completed in approximately fifteen minutes, with an effective logical error rate one tenth the physical rate. That is a number a reviewer can hold. It places the Quantum X announcement in a context: decoders must now operate on boards that carry such quantities of entangled work, and their speed and accuracy become the true constraint. The date is not incidental. The company says its model was trained on synthetic samples only and not on real hardware shots from the Google dataset. If a decoder trained on invented noise can beat decoders fitted to real noise, the margin should be published; a benchmark without a margin is a proof of hope. The pattern is familiar: the gap between stated readiness and observed readiness remains instructive. The registry of such gaps grows longer than the number of published margins. Until the firms that make these claims open their logs to comparable measurement, the number that matters most will remain the one they decline to print. Let that stand as the ledger for this morning. The time has not passed for replication, and Prof. Sharon says the next objective is to extend the result across device centres, but the ordinary reader might ask for a foot of the calculation rather than a promise of more. LONDON, 21 AUGUST — The claim that a quantum computer has surpassed classical methods is no longer extraordinary. What remains extraordinary is the proof. In July, IBM and the University of Chicago reported a seventy-logical-qubit computation that completed in fifteen minutes and, they argued, lies beyond practical classical simulation. The result came with a statistical guarantee of fidelity. But the guarantee is a lower bound, not a certificate, and the field's own practitioners now speak of verification as a process without end. Jay Gambetta, IBM's research director, told reporters the demonstrations prove "that quantum computers can solve problems that go beyond the reach of classical methods that could run on the biggest classical computers." He called it a foundation for trusting quantum systems as they scale. The claim was for three papers, all yet to undergo peer review. Each approached validation from a different angle: one using random circuits with error detection, two using error mitigation on physical simulations. The Chicago collaboration built circuits entirely of Clifford gates, which are classically simulatable, and then inserted non-Clifford gates in positions that preserved error detection. The approach, called spacetime code, allowed the team to discard failed runs and still achieve a 95 percent confidence guarantee on the fidelity of the surviving ones. Bill Fefferman, a professor at Chicago, said the experiment "develops techniques to better characterize the fidelity of hard quantum states under noise." His student Soumik Ghosh added that such advances "have the potential to unlock practical applications for the next generation of quantum computers." But the other two papers, from the startups Qedma and Algorithmiq, leaned on error mitigation rather than error detection. They simulated a model of a magnetic material under a regular pulse and compared results with leading classical methods on Japan's Fugaku supercomputer and an Nvidia H100 server. At 35 qubits all agreed. At 51 qubits the classical methods broke down. At 74 they were out of reach. The authors did not claim quantum advantage; they claimed consistency. Netanel Lindner, Qedma's chief technology officer, said the team ran the same simulation on Quantinuum's and Helios's trapped-ion machines and observed the same behaviour: "The results we get with Quantinuum...are in perfect match with what we got from IBM." Algorithmiq's CEO, Sabrina Maniscalco, called the stability across hardware and noise levels evidence rather than coincidence. Dominik Hangleiter, a postdoctoral researcher at ETH Zurich, reviewed the work and said both papers seemed to show the problem was hard to simulate classically, which he judged good science. "They shouldn't make much stronger claims than that," he said. Jens Eisert of the Free University of Berlin put it in longer form: verification "is not a single procedure, but rather a process of building confidence through a portfolio of complementary validation methods." That is the quietest way to say the obvious: a portfolio of methods is a collection, not a proof. The scramble to trust these machines comes at a moment when the hardware itself is improving faster than the methods to check it. In late December, Craig Gidney, a Google researcher, published a log of logical qubit half-lives under repetition codes: one hundred microseconds in 2014, three milliseconds in 2021, three hundred milliseconds in 2023, and two hours in 2024. "It took nearly a decade for the half life to approach one second, and then one year later it's suddenly measured in hours?" he wrote. He attributed the leap to gap engineering that mitigated cosmic rays. But he also warned of further hurdles, and his model predicts a superexponential lull followed by a sudden jump. The same pattern, he said, will soon apply to full error correction. That is the operating environment. Decoders, the classical systems that read errors and decide corrections, are now the bottleneck. Quantum X Labs in Tel Aviv announced yesterday that its AI-driven decoder, trained only on synthetic data, outperformed Google's published matching-family benchmarks on a real-hardware surface-code dataset. Prof. Nir Sharon, the company's chief quantum scientist, called the result "a meaningful validation point" while adding that it is one configuration and replication remains the objective. He did not publish the margin. The gap between stated readiness and observed readiness remains instructive, as usual. What the outside observer can hold onto is the number of verbs. IBM says its result "demonstrates" advantage. Gidney says the quality barrier "will fall within the next five years." Qedma says the oscillations "exist." A Victorian reader would ask for the ledger, not the testimony. The ledger exists: seventy logical qubits, 2,415 logical two-qubit operations, 468 logical T gates, an effective logical error rate one tenth of the physical rate, and a confidence interval that spans 95 percent. Those are figures a reviewer can carry. The smaller figures, the ones not printed, are the ones that decide whether the claim holds. For the record, then, the field's own language admits the uncertainty. Hangleiter says the authors should not claim more than that the problem seems hard to simulate. Eisert says confidence is built across methods. Gambetta says the demonstrations are proof of scaling. The words differ. The underlying shape is the same: a decade of progress compressed into a year, and verification still running behind the machine it checks. The scramble is to keep the ledger accurate while the entries change daily. That work continues, without ceremony, as all such work does. The ledger of quantum computation is older than the machines it polices. On its first page sits Peter Shor's 1994 paper, which showed that a quantum computer could factor large numbers with a speed that classical machines would never match. That page has never been closed. For three decades, the same entry recurs: a claim of computational superiority, followed by months or years of argument about what the result actually proves. The entry for 2026 is no different, save for the confidence interval printed beside it. The question is not whether the ledger grows, but how long it has been growing, and why the verification column trails the announcement column by so wide a margin. It has been true since the beginning. In 2019, Google announced that its Sycamore processor had performed a random circuit sampling task beyond the reach of the best classical supercomputers. IBM immediately contested the claim, noting that a classical simulation could be done in days rather than the ten thousand years Google projected. The episode established a pattern that has not since broken: a hardware demonstration, a classical rebuttal, a revised estimate. Each cycle takes months, and by the time the numbers are settled, the hardware has moved on to a larger circuit. The pattern has held for seven years, and its persistence is itself a finding. IBM's July announcement follows the same geometry. With the University of Chicago, the company reported a seventy-qubit logical computation completed in fifteen minutes, on a spacetimes code that allowed error detection and a 95 percent confidence guarantee on the fidelity of surviving runs. The result, as reported in the arXiv preprint, used 2,415 logical two-qubit operations and 468 logical T gates, with an effective logical error rate one tenth of the physical rate. Jay Gambetta, IBM's research director, called it a demonstration beyond the practical reach of classical computers. But the confidence guarantee is a lower bound, not a certificate. It tells you how faithfully the circuit ran on the runs that were not discarded; it does not tell you whether the computation was useful. Dominik Hangleiter, a postdoctoral researcher at ETH Zurich, said the papers should not claim more than that the problem seems hard to simulate. Jens Eisert of the Free University of Berlin described verification as “a process of building confidence through a portfolio of complementary validation methods.” A portfolio is a collection, not a proof. The hardware has been improving faster than the methods to check it. Craig Gidney, writing on 24 December, published a log of logical qubit half-lives under repetition codes: one hundred microseconds in 2014 with nine qubits, three milliseconds in 2021 with twenty-one, three hundred milliseconds in 2023 with fifty-one, and two hours in 2024 with fifty-nine. The jump from three hundred milliseconds to two hours in a single year came from gap engineering that mitigated cosmic rays. Gidney called it a lull followed by a FOOM, and he predicted the quality barrier would fall within five years. But that prediction is itself a hypothesis, not a measurement. The gap between stated readiness and observed readiness remains instructive, as usual. The scramble to trust these machines is not new. It began with the first quantum algorithms and has never stopped. What is new is the rate of change. The IBM claim, the Quantum X Labs decoder result, and Gidney's timeline all point to a field that is compressing decades of work into months. But the verification tools are still calibrated for the old pace. The 2026 entry in the ledger will be revised, as all entries are revised, and the revision will take months. By the time it is settled, the hardware will have moved on. For the record, then, the question of how long this has been true has a definite answer. It has been true since 1994, and it will be true for some time yet. The ledger does not close; it accumulates. The pattern is as stable as the equations it polices. The only change is the size of the numbers printed beside the dates. The pattern predates the equations. In 1858, when the first transatlantic cable was laid, the world celebrated a triumph. Queen Victoria exchanged telegrams with President Buchanan, and the press pronounced the ocean a parchment. But within three weeks, the signal withered into silence. The engineers had overclaimed; the insulation had failed under pressure, and the operators had no way to verify whether the messages they sent were the messages received. A subsequent inquiry revealed a gap between promise and performance, measured but unheeded, and it took years and a second cable before transatlantic telegraphy became dependable. As Tom Standage documents in *The Victorian Internet*, the scramble to trust the new medium was not a matter of faith but of repeated, patient calibration against the physical world. That geometry recurs this morning. IBM announces a seventy-logical-qubit computation that 'demonstrates' quantum advantage, with a 95 percent confidence guarantee on the fidelity of surviving runs. But the guarantee is a lower bound, not a certificate; the discarded runs are not printed, and the verification method itself is a live project. Quantum X Labs reports a decoder trained on synthetic data that outperforms Google's matching-family benchmarks, yet withholds the margin. The ledger grows, but the verification column trails, as it always trails when the tool outruns its inspector. Who is exposed? The operator who must act on an unverified result. In 1858, it was the telegraph clerk who pressed the key and trusted the copper. Today, it is the systems engineer who must migrate a financial network to post-quantum cryptography, or the HPC centre that must schedule a workload on an unproven processor. They are the ones who break if the claim is false. How long has this been true? Since the first time a machine produced a result its maker could not independently check. The cable failed in 1858; the verification of that failure took months. The quantum advantage claim of 2026 will be settled in years, by which point the hardware will have moved on. This is not a novel crisis. It is the ordinary condition of engineering: the tool advances, and the measure limps behind. The scramble is to keep the ledger honest while the entries change daily, and that work has been underway for as long as there have been cables to lay and circuits to trust. Meanwhile, the migration of cryptographic infrastructure to post-quantum algorithms proceeds without a schedule. The Vancouver concern Quantum Secure Encryption announced in December the release of its qREK software development kit, a tool for local key generation backed by quantum entropy, compatible with AES, RSA, and the NIST-recommended Kyber and Dilithium algorithms (QSE press release, 9 December 2025). The announcement records a tool, not a timetable. No enterprise, utility, or government has yet published a date by which its systems will be rotated onto the new standards. The gap between the published threat and the planned response stretches wider each quarter. For the record, then: the migration has not been scheduled. The parties exposed are those who hold secrets of long duration: the operators of power grids, hospitals, and banks, who must protect data that will still be sensitive decades hence. The threat has been known since 1994, and the migration has been urged for years. Yet no binding date has been set. The logs will reflect what the summaries omit. Who is exposed? The operator who must act on an unverified result. In 1858 it was the telegraph clerk who pressed the key and trusted the copper; today it is the systems engineer who must migrate a financial network to post-quantum cryptography, or the HPC centre that must schedule a workload on an unproven processor. The parties are concrete. The EuroQCI project, which proposes to shield power grids and hospitals with fibre and satellite QKD, is one such party; the enterprises adopting qREK to generate keys locally are another. The HPC centres integrating quantum hardware through QDMI, as described in the recent IQM case study, are a third. None has yet published a date for the rotation of its certificates, and the gap between the published threat and the planned response stretches wider each quarter. For the record, then: the migration has not been scheduled, and the exposed are those who must act before the numbers are settled. —Inspector Grey Dispatch from The Prepared E0