The Picture of Ninety-Four Qubits
On error detection, postselection, and the portrait that ages in the attic
Keywords: quantum error correction; logical qubits; postselection; quantum error detection; blinding; scientific fraud; Helios; Iceberg code; benchmarking; research integrity
“To lose one experimental run, Professor, may be regarded as a misfortune; to lose ninety-six point eight per cent of them looks like carelessness.”
— Lady Bracknell, had she read the supplementary information
I. In which a number becomes a work of art
There is only one thing in the world worse than being talked about as a quantum computer, and that is being talked about as a quantum computer with ninety-four logical qubits. The number has everything a number could wish for: it is large enough to be impressive, odd enough to be memorable, and precise enough to suggest that someone, somewhere, measured it. It has been launched, press-released, and blogged. It has dined out all season. Like most celebrated beauties, it owes a great deal to lighting.
The lighting, in this case, is a pair of technologies with names so similar that only a pedant — or a prosecutor — would insist upon the difference. One is quantum error correction, the physician who diagnoses the disease and cures it while the patient continues about his business. The other is quantum error detection, the mirror that merely tells the patient he is unwell, after which the patient is quietly removed from the guest list. Both produce spotless dinner parties. Only one produces a healthy society.
I have been reading, with the attention one usually reserves for the society pages, the technical record of the machine behind the famous number: Quantinuum’s Helios, a ninety-eight-qubit trapped-ion device of genuine and considerable engineering achievement (Ransford et al., 2025), and the two experimental papers that report its widest encoded results (Dasu et al., 2026; Perlin et al., 2026). Permit me to be clear at the outset, for clarity is the only courtesy one owes both the accused and the audience: the hardware is real, the encoding is real, and several of the results are admirable. It is not the experiment that concerns me. It is the noun.
II. The meaning of “logical,” a word that has suffered much
In coding theory, a quantum error-correcting code carries three parameters, conventionally written [[n, k, d]]: n physical qubits encoding k logical degrees of freedom, with distance d. The parameter k is a theorem. It says that the code space, as a matter of algebra, has dimension 2^k. The Helios experiments implemented an Iceberg code of the family [[k+2, k, 2]], and at its largest block a [[96, 94, 2]] encoding — a rate of nearly 0.98, which is to say ninety-four logical degrees of freedom packed into ninety-six physical qubits with only two left over to mind the door (Dasu et al., 2026). This is a genuine construction, and calling those ninety-four factors “logical qubits” is, as nomenclature, perfectly respectable.
But nomenclature is where the respectability ends, and where the evening’s entertainment begins. A code parameter tells you the size of the room; it tells you nothing about whether anyone in it is dancing. The distance of the [[96, 94, 2]] code is two, and a distance-two code corrects, for arbitrary errors at unknown locations, exactly nothing. The correction radius of a distance-d code is ⌊(d−1)/2⌋; set d = 2 and the radius is zero (Knill & Laflamme, 1997; Terhal, 2015). The code detects a single error. It announces that something has gone wrong — and then the run in which it went wrong is discarded.
This is the entire mechanism, and one must admire its elegance. One does not correct the failure; one declines to count it. The machine is then said to have performed with high fidelity, and so it has — among the runs it consented to acknowledge. We are all in the noise, but some of us are postselecting the stars.
III. The portrait in the attic
Dorian Gray, you will recall, remained forever young because the consequences of his life were transferred to a portrait locked upstairs. Postselection is the same arrangement, with better tailors. The submitted runs are the life; the accepted runs are the face; the discarded runs are the portrait, and it is in the attic, and it is hideous, and it is the truth.
The supplementary tables of the 94-factor experiment supply the ledger, as supplementary tables so obligingly do. The flagship result is a 94-factor logical GHZ state — a wide and genuinely non-trivial preparation, a real test of long-range correlation and syndrome handling. Without leakage heralding, the encoded conditional fidelity is 0.87 and the overall acceptance 0.36. With leakage heralding — that is, with a second opportunity to refuse the unflattering runs — the conditional fidelity rises to a radiant 0.949, while the acceptance falls to 0.25 (Dasu et al., 2026). Read the pair of numbers as one must always read pairs of numbers in this literature: three runs in four were burned to keep the complexion of the fourth.
The 64-logical-degree-of-freedom experiment — a hardware simulation of a three-dimensional XY model, again real physics on real hardware — makes the trade visible as a curve, and I commend it to you as Figure 1. At two Trotter steps, the leakage-heralded encoded fidelity is 0.69 and 23.2% of submitted shots survive. At ten Trotter steps, the encoded conditional fidelity still holds a point estimate above its unencoded comparator — about 3.3 times higher — but the acceptance has collapsed to 3.2%: ninety-six point eight per cent of everything the machine did was discarded to obtain it (Dasu et al., 2026). The deeper the computation, the more of reality must be refused to keep the picture pretty.
Figure 1 :
There is a deeper discipline in the paper, and honesty requires me to praise it before I return to my cynicism — a cynic, after all, being merely a man who knows the width of everything and the depth of nothing, and I aspire to know both. A concatenated construction, I₈∘I₆, reaches distance four over forty-eight logical degrees of freedom, and distance four corrects one arbitrary unlocated error. Here, at last, is a correction-capable block. And here the record is instructive: the quoted QEC acceptance of 0.62 is a stage-conditional figure, not an end-to-end one; of 5,000 submitted shots in each of two bases, 708 and 958 survived every filter — end-to-end retentions of 14.2% and 19.2%. Among those survivors, zero logical failures were observed, which exact binomial arithmetic converts into honest upper bounds on the conditional failure probability (Dasu et al., 2026). Zero failures in the accepted shots is a real and encouraging result. It is also, definitionally, a statement about the shots that were allowed to testify.
The companion paper does something harder and rarer. Using the Steane [[7, 1, 3]] code, Perlin et al. (2026) compiled algorithms to a protected Clifford+T gate set with dynamic feedback and repeat-until-success preparation, and — in the HHL configurations with two and four actively inserted error-correction cycles — performed correction during the workload rather than around it. This is the genuine article: entropy removed mid-circuit, in one integrated hardware process, at small scale and near break-even. Their largest 12-logical-qubit QAOA circuits, by contrast, insert zero correction cycles into the algorithm body, retain 349 of 500 shots, and beat the random-guess null of 1/4096 with a success probability of 0.034 — which is 16% of the noise-free reference value, against 84.7% for the unencoded circuit on the same device (Perlin et al., 2026). The authors themselves treat “better than random” as the low bar it is. I merely note how many headlines would not.
IV. What every other science calls this
And now to the matter that should end the conversation, and strangely never does.
In every other empirical discipline, the procedure I have just described has a name, and the name is not “quantum error detection.” The name is selective reporting of favourable outcomes, and it is misconduct.
Consider medicine. A trial administers a drug to five thousand patients, observes the outcomes, discards the eighty-six per cent in whom the treatment failed or the data proved inconvenient, and publishes the conditional recovery rate among the remainder as evidence of efficacy. No regulator would call this medicine; no journal would call it science; the courts, in due course, would call it something shorter. The entire edifice of the randomised controlled trial — the intention-to-treat principle, which insists that every randomised patient be counted whether or not the therapy flattered them — exists precisely to forbid the quantum community’s favourite denominator. Medicine learned, at the cost of actual lives, that the investigator who chooses his witnesses has chosen his verdict.
Consider psychology, a field that discovered its own replication crisis in the way Dorian discovered his portrait — suddenly, publicly, and after years of assurances that everything was fine. Its remedy was preregistration: the hypothesis, the measure, the analysis and the exclusions fixed before the data are seen, so that the researcher may not wander the garden of forking paths picking the route that ends at a significant result. The sin that preregistration exists to prevent is the sin of deciding, after the fact, which observations count.
Consider particle physics, a discipline not otherwise noted for humility, which nonetheless practices blind analysis: the signal region of the data is hidden from the experimenters until every cut, every calibration and every selection rule is frozen. The physicist is not permitted to see which events survive before committing to the filter, because the community understood long ago that a selection rule tuned on the answer is not a measurement but a memoir.
The pattern is universal enough to deserve an epigram: science is the art of not being allowed to choose your own evidence. A blinded test — one in which the researchers do not select which runs, which patients, which events or which outcomes will testify — is not a gold standard, a nicety, or a bureaucratic ritual. It is the defining act. It is the moment the investigator surrenders the power to edit reality and thereby earns the right to report it. Everything else is criticism; and criticism, as I have said elsewhere, is the only civilised form of autobiography.
Postselection as practised in the wide logical-qubit demonstrations inverts this act. The detection circuit looks at each run, pronounces upon it, and the failed runs are removed from the ledger before the fidelity is computed. The conditional number that survives is then quoted beside the noun “logical qubits,” and the noun does the work the experiment did not: it transfers, by pure association, the aura of fault tolerance — the property of sustaining arbitrarily long protected computation, whose entire theoretical edifice from Shor (1996) through the threshold theorems (Aharonov & Ben-Or, 2008; Knill, 2005) concerns errors corrected, not errors refused — onto a result that is, mechanically, a count of invitations accepted. In any other science this transfer would be called fraud. In quantum computing it is called Tuesday, and on Tuesdays the press releases go out. One launch page asserted that the machine was used to “make 94 logical qubits” with no qualifier attached; a companion commercial release, to its credit, restored the words “error detected” for the 94 and reserved “error corrected” for the 48 (Quantinuum, 2025). The qualifiers exist. They simply attend the wrong parties.
Let me be scrupulously fair, because fairness is the last refuge of the devastating. Postselection is not always a vice. Error detection with heralded rejection is a legitimate and powerful tool — for state preparation, for benchmarking syndrome hardware, for erasure-aware schemes in which the location of the error is genuinely known: Zhang et al. (2026) demonstrate on hardware that even a distance-two [[4, 2, 2]] code corrects a located erasure, which is correction in earnest and not a euphemism. The vice is not the filter. The vice is the invoice: presenting a conditional, discard-weighted number as an unconditional measure of computational capability, and letting the width of the code stand in for the depth of the evidence.
V. The honest exception, and why it matters
The strongest evidence against my indictment comes, deliciously, from within the field itself. Bluvstein et al. (2026), working with reconfigurable neutral-atom arrays, report that increasing code distance from three to five lowers the logical error per round by a factor of 2.14(13) — and state explicitly that no postselection is used for that comparison. Every run testifies. The improvement is an improvement of the whole population, not of the survivors. That single sentence — no postselection is used — is worth more than any width number published to date, because it is the sound of a result that does not fear its own denominator. The same programme demonstrates repeated correction, logical teleportation, universal non-Clifford ingredients and deep logical circuits across distinct experimental objects, with the postselection used in the deep-circuit sections disclosed as such in the methods (Bluvstein et al., 2026). Here is the field’s own proof that it knows the difference between a measurement and a memoir: when the claim is about scaling, the filter is removed; when the filter is present, it is confessed.
Which raises the obvious question: if the honest design is possible, why is it not the standard? The answer is that honesty is expensive in precisely the currency these demonstrations are minted to display. Require every submitted shot to count, and the headline fidelity falls. Require the acceptance rate beside the fidelity, and the headline is qualified. Require a scaling grid — several progressively stronger protection configurations, each run at several workload sizes, with loss, acceptance and time-to-solution recorded for every cell, every configuration present at every workload point, and success defined in advance rather than anointed afterwards — and no published experiment, on any platform, yet qualifies. The strongest distance-scaling results in the literature supply two configurations at one workload each, where the honest design demands at least three configurations across at least two workloads. This is not a criticism of those experiments; it is the specification of an experiment that has not yet been performed. The grid is empty, and the emptiness is the finding.
VI. A modest proposal, seriously offered
Wilde is supposed to have said that he put his talent into his work and his genius into his life; the quantum computing industry has reversed the apportionment. The remedy, however, is neither talent nor genius but bookkeeping. Every public claim of a logical-qubit count should carry, as a matter of course, the minimum record that would allow a sceptic to falsify it: the width actually exercised; the protected depth; a block-level or task-level failure metric, not merely a per-factor average; the acceptance or restart policy and the measured acceptance, end to end, submitted shots in the denominator; the runtime or time-to-solution; and a protection descriptor naming the code, the error model, the detection-versus-correction status, and the cadence of any active correction. Six fields. No more than the label on a decent bottle of wine, and for the same reason: so the buyer knows what he is drinking.
Beyond disclosure lies the harder demand: claims should be tested against benchmarks fixed independently of the claimant, under decision rules stated before the data are seen, with the selection of runs, patients, events or cells removed from the investigators’ hands. This is not a radical programme. It is Tuesday in every other science. The day the quantum field adopts it without being asked is the day it stops needing auditors and starts deserving its press releases.
VII. Epilogue, with epigrams
The 94 logical qubits are real, in the sense that a mirror is real: they show you exactly what is placed before them and nothing about what was cropped from the frame. Encoded width is a legitimate property of an implemented code; it is not a sufficient statistic for a functioning protected processor, and a community that treats it as one has not built a computer so much as a mirror with a marketing department.
The experience of reading this literature has been, as all experience is, the name we give to our mistakes — in this case, the mistake of believing that the noun “logical” carried the adjective “corrected” as a dowry. It does not. Between detection and correction lies the whole difference between vanity and virtue: the one arranges the evidence, the other survives it.
And so I end where every consideration of modern art must end, with the portrait. Somewhere in an attic of supplementary tables lies the true picture of the ninety-four qubits: 5,000 submitted runs, four-fifths of them disfigured by noise, stacked like rejected canvases against the wall. The face in the drawing room is beautiful, and it is honest work, and it deserves its frame. But in matters of grave importance, sincerity — not style — is the vital thing; and the sincerity of an experiment lives, as it always has, in the runs it was willing to count.
References
Aharonov, D., & Ben-Or, M. (2008). Fault-tolerant quantum computation with constant error rate. SIAM Journal on Computing, 38(4), 1207–1282. https://doi.org/10.1137/S0097539799359385
Bluvstein, D., Geim, A. A., Li, S. H., et al. (2026). A fault-tolerant neutral-atom architecture for universal quantum computation. Nature, 649, 39–46. https://doi.org/10.1038/s41586-025-09848-5
Campbell, E. T., Terhal, B. M., & Vuillot, C. (2017). Roads towards fault-tolerant universal quantum computation. Nature, 549, 172–179. https://doi.org/10.1038/nature23460
Dasu, S., DeCross, M., Guo, A. Y., et al. (2026). Computing with many encoded logical qubits beyond break-even. arXiv. https://doi.org/10.48550/arXiv.2602.22211
Fowler, A. G., Mariantoni, M., Martinis, J. M., & Cleland, A. N. (2012). Surface codes: Towards practical large-scale quantum computation. Physical Review A, 86, 032324. https://doi.org/10.1103/PhysRevA.86.032324
Google Quantum AI. (2023). Suppressing quantum errors by scaling a surface code logical qubit. Nature, 614, 676–681. https://doi.org/10.1038/s41586-022-05434-1
Google Quantum AI and Collaborators. (2025). Quantum error correction below the surface code threshold. Nature, 638, 920–926. https://doi.org/10.1038/s41586-024-08449-y
Knill, E. (2005). Quantum computing with realistically noisy devices. Nature, 434, 39–44. https://doi.org/10.1038/nature03350
Knill, E., & Laflamme, R. (1997). Theory of quantum error-correcting codes. Physical Review A, 55(2), 900–911. https://doi.org/10.1103/PhysRevA.55.900
Perlin, M. A., He, Z., Armenakas, A. A., et al. (2026). Fault-tolerant execution of error-corrected quantum algorithms. arXiv. https://doi.org/10.48550/arXiv.2603.04584
Proctor, T., Rudinger, K., Young, K., Nielsen, E., & Blume-Kohout, R. (2022). Measuring the capabilities of quantum computers. Nature Physics, 18, 75–79. https://doi.org/10.1038/s41567-021-01409-7
Quantinuum. (2025, November 5). Introducing Helios: The most accurate quantum computer in the world. https://www.quantinuum.com/blog/introducing-helios-the-most-accurate-quantum-computer-in-the-world
Quantinuum. (2026, March 4). Skinny logic: Quantum codes go on a diet. https://www.quantinuum.com/blog/skinny-logic-quantum-codes-go-on-a-diet
Ransford, A., Allman, M. S., Arkinstall, J., et al. (2025). Helios: A 98-qubit trapped-ion quantum computer. arXiv. https://arxiv.org/abs/2511.05465
Shor, P. W. (1996). Fault-tolerant quantum computation. In Proceedings of the 37th Annual Symposium on Foundations of Computer Science (pp. 56–65). IEEE. https://doi.org/10.1109/SFCS.1996.548464
Steane, A. M. (1996). Error correcting codes in quantum theory. Physical Review Letters, 77(5), 793–797. https://doi.org/10.1103/PhysRevLett.77.793
Terhal, B. M. (2015). Quantum error correction for quantum memories. Reviews of Modern Physics, 87(2), 307–346. https://doi.org/10.1103/RevModPhys.87.307
Zhang, B., Liu, G., Bornet, G., et al. (2026). Logical qubits with erasure conversion using metastable neutral atoms. Nature Physics, 22, 910–916. https://doi.org/10.1038/s41567-026-03309-0