Quantinuum's C4-Helix Code on Helios, 2 September 2026: Two Logical Qubits in Twenty Ions, the Full Clifford Group, and What the Paper Says It Has Not Done
In April 1950 Richard Hamming published "Error Detecting and Error Correcting Codes" in the Bell System Technical Journal. His seven-bit code corrected one error in every protected word, and the paper set out the general problem of adding enough redundancy to a message that a fault leaves a trace. Computing on the protected words without first unpacking them, so that the protection never lapses, is the harder half of the same problem, and it is the half a Quantinuum team reported on this week.
On 2 September 2026 fourteen authors from Quantinuum's offices in Broomfield, Colorado and London posted "Experimental validation of a compact fault-tolerant architecture for trapped ions", arXiv:2609.03194. The paper's own framing is careful: quantum error correction has begun to deliver logical operations that beat their unencoded physical counterparts, and useful fault-tolerant computation will require more than low-error quantum memory. The team then presents one specific code, the C4-Helix code, and reports three experiments on the company's 98-qubit Helios processor that together, in the authors' words, establish it as a hardware-validated fault-tolerant architecture and no longer solely a quantum memory. This post reads what was measured, what the numbers mean against the processor's own baseline, and what the paper says still stands between this result and a computation of practical value.
Practical takeaway. A logical error rate is only meaningful next to three other numbers: the physical error rate on the same machine, the number of physical qubits the code consumes per logical qubit, and the set of operations the code can perform while staying protected. This paper supplies all four, which is rarer than it should be, and it says in its own words that universal computation is not yet demonstrated.
What the [[20,2,6]] C4-Helix code is, and why twenty qubits for two matters
The code notation gives the whole shape. Twenty physical qubits encode two logical qubits at distance six, meaning the smallest error that changes a logical value without leaving a detectable trace touches six physical qubits. The construction concatenates a ten-qubit twisted toric code, written [[10,2,3]], with the four-qubit C4 code, written [[4,2,2]]. Trapped-ion processors can move ions physically, so any qubit can interact with any other; the authors use that freedom to pick a code with long-range connections, which is what lets twenty qubits carry two logical qubits at this distance. The paper states the code needs roughly 3.5 times fewer physical qubits than a rotated surface code of comparable distance, the code family that superconducting processors, whose qubits sit fixed on a chip, are built around.
Qubit count is one axis. The other is what the code can do without leaving protection. Most of the logical Clifford operations on C4-Helix run as depth-one transversal gates or as automorphisms, which amount to single-qubit gates plus a relabeling of which ion is which, and a depth-four fold-transversal S gate completes the set. That is why the authors call it an architecture: the encoding, the operations and the interface to other codes are designed together. Quentir's earlier reading of the Bharti, Haug and Tanggara proof that fault tolerance carries an unavoidable logarithmic overhead explains why the size of that overhead, and not its existence, is the design variable that separates codes.
The processor underneath: Helios, 98 barium ions, described on 7 November 2025
Helios is Quantinuum's trapped-ion machine, described in the company's own technical paper of 7 November 2025, arXiv:2511.05465, by 186 authors. It holds 98 qubits encoded in the hyperfine states of barium-137 ions, arranged in a quantum charge-coupled device with a rotatable storage ring feeding two operation regions through a junction. The reported physical numbers are single-qubit gate infidelity of 2.5 × 10−5, two-qubit gate infidelity of 7.9 × 10−4, and state preparation and measurement error of 4.8 × 10−4. Those figures are the baseline every logical result in the new paper is compared against, and the comparison is the only reason the logical numbers mean anything. The junction that connects Helios's regions is a mechanical one that moves ions; it is a different object from the Josephson junction inside a superconducting qubit that Princeton's new quantum institute is aimed at, and the two platforms carry different error budgets for that reason.
Three experiments, all reported without postselection
The first experiment is memory. The team prepared the two logical qubits, ran twenty rounds of syndrome extraction with leakage repumping and circuit-level leakage reduction, then measured and decoded offline with Quantinuum's Frontier decoder. The measured logical error rate was 4.6 × 10−5 per logical qubit per correction cycle for the [[20,2,6]] code, with a 95 percent confidence interval running from 2.0 × 10−5 to 1.08 × 10−4. The smaller [[10,2,3]] code on the same machine gave 2.1 × 10−4. Concatenating with C4 therefore cut the per-cycle error by roughly a factor of four or five, which is the behavior a working code is supposed to show as its distance grows.
The second experiment is computation. The team ran two-qubit randomized benchmarking of the complete Clifford group on the two logical qubits of one code block, with active syndrome extraction interleaved between gates, and measured 2.8 × 10−4 error per two-qubit logical Clifford, against 1.2 × 10−3 for the same benchmark on unencoded physical qubits. The paper also reports that adaptive syndrome extraction cut the physical two-qubit gate count for the logical S gates by 33 percent and the wall-clock shot time by 23 percent. A logical Clifford with about a quarter of the physical version's error is the number a buyer should carry away, because it is the result that measures operations, where the earlier one measured storage.
The third experiment is the interface. Using what the authors call a chain-map CNOT, they entangled the two C4-Helix logical qubits with one logical qubit held in a 25-qubit distance-five rotated surface code and prepared a three-qubit GHZ state across the two code families. The fidelity lower bound was 99.925 percent, with a confidence interval from 99.68 percent to 99.99 percent, against 99.54 percent for the physical version. This experiment exists for one reason the paper states plainly: C4-Helix does not itself produce the non-Clifford resource states, usually called magic states, that universal computation requires, and the architecture plans to import them from a surface or color code block through exactly this kind of interface.
All three results were obtained without postselection, meaning no failed runs were discarded to improve the average, and the paper states that in each case the encoded implementation outperformed its unencoded physical counterpart; for the memory experiment the comparison printed beside the headline figure is with the smaller [[10,2,3]] logical code, and the physical memory baseline is not among the numbers this post can quote. The paper includes a figure showing what postselection on decoder confidence would do, and separately reports that a decoder-confidence threshold of 3.9 nats would have removed every observed logical failure in the C4-Helix memory dataset, but the headline figures do not rely on it. That distinction matters for anyone comparing this paper to results elsewhere, since several published logical-qubit numbers in this field depend on discarding a fraction of shots, and the two kinds of figure cannot be compared directly.
What the paper says it has not done
The discussion section is explicit. The experiment does not implement a universal Clifford+T computation. It establishes the Clifford substrate and the heterogeneous-code interface needed to construct one. The logical error rates achieved, around 10−4 per correction cycle and per logical Clifford, sit two to four orders of magnitude above the 10−6 to 10−8 range that the authors themselves define as the early fault-tolerant regime, the range they associate with scientifically and commercially valuable computations beyond the reach of classical supercomputers. Reaching that range, on the paper's circuit-level simulations, requires physical infidelities improved into the 10−4 regime, roughly an order of magnitude better than Helios's current two-qubit gates, and the authors note that trapped-ion testbeds have already reported gates at that level. They also sketch larger relatives of the code, a [[60,2,12]] Carbon-Helix and a [[100,2,18]] Double-Helix, whose circuit-level performance is left to future work.
Two further limits are worth naming because the paper does not hide them. Decoding was done offline for the memory experiment, so the real-time decoding a running computation would need is not part of this result. And the uncertainty on the memory figure is wide: the upper end of the confidence interval on the [[20,2,6]] error is more than double the point estimate, because few logical failures were observed in the dataset behind it. The claim that survives all of this is specific: on Helios, in September 2026, a compact code kept two logical qubits, computed the whole Clifford group on them, and handed them to a different code; the logical Clifford and GHZ results beat their stated physical baselines, and the memory result improved on the smaller logical code.
How Quentir Reads It
The procurement reading follows from the chronology. On 7 November 2025 Quantinuum published Helios's physical error rates. On 2 September 2026 the same company published logical error rates on that machine with the physical baseline printed beside them, the qubit overhead stated, and the missing piece, universal computation, named in its own discussion section. Quentir's reading of Roeland Wiersema's classical simulation of D-Wave's advantage experiment argued that a vendor claim becomes checkable when the vendor publishes enough to be tested; this paper is the form that takes for error correction, and it should be held to the same standard when the next result on a bigger code arrives.
For a research director or a program office the useful question is which of the four numbers in this paper a proposal in front of them can produce for its own hardware: physical error rate, logical error rate on the same machine, physical qubits per logical qubit, and the set of protected operations. A proposal should report all four before its readiness can be assessed, and should separately state how non-Clifford resources will enter the protected computation; with those in hand it can be placed on a readiness ladder with a date attached. This site's Medicine Monitor asked the same question of a clinical result when it examined when the error check happens and what it buys; here the check happens between every logical gate, and what it bought was a measured error per two-qubit Clifford of 2.8 × 10−4 logically against 1.2 × 10−3 physically.
The item to watch is dated by the paper itself: a Helios physical two-qubit infidelity in the 10−4 range, which on the authors' simulations would move this same code into the regime they associate with computations beyond classical reach. Quentir's Signature Report No. 3, Quantum-AI Convergence 2026, keeps the published logical-qubit results of the past two years side by side with their physical baselines, overheads and postselection status, so a committee can see which vendors have supplied all four numbers and which have supplied one. The public analysis on this site continues daily at quentir.ai/blog.
Sources: Noah Berthusen, Ali Lavasani, Asmae Benhemou, M. S. Allman, Joan Dreiling, Brian Estey, Cameron Foltz, Trent Jacobs, Michael Mills, Annie Jihyun Park, Adam P. Reed, David Hayes, Tzvetan S. Metodi and Andrew C. Potter, "Experimental validation of a compact fault-tolerant architecture for trapped ions", arXiv:2609.03194, Quantinuum, Broomfield and London, 2 September 2026, read in full this run including the discussion, the acknowledgments and the appendices, for the [[20,2,6]] C4-Helix construction from the [[10,2,3]] twisted toric code and the [[4,2,2]] C4 code, the roughly 3.5-fold spatial saving against rotated surface codes of comparable distance, the twenty rounds of syndrome extraction with leakage repumping and leakage reduction units, the per-cycle logical error of 4.6 × 10−5 with its 95 percent Wilson interval and the 2.1 × 10−4 figure for the smaller code, the two-qubit logical Clifford error of 2.8 × 10−4 against the physical 1.2 × 10−3, the 33 percent gate-count and 23 percent wall-clock reductions from adaptive syndrome extraction, the chain-map CNOT to the [[25,1,5]] rotated surface code and the GHZ fidelity bounds of 99.925 and 99.54 percent, the absence of postselection in the headline figures and the 3.9-nat gap that would remove every logical failure, offline decoding with the Frontier decoder, the statement that no universal Clifford+T computation was implemented, the 10−6 to 10−8 target regime and the circuit-level simulations placing it at physical infidelities in the 10−4 range, and the [[60,2,12]] and [[100,2,18]] extensions. Anthony Ransford, M. S. Allman, Jake Arkinstall and 183 co-authors, "Helios: A 98-qubit trapped-ion quantum computer", arXiv:2511.05465, Quantinuum, 7 November 2025, for the 98 barium-137 hyperfine qubits, the rotatable storage ring and junction architecture, and the physical infidelities of 2.5 × 10−5 for single-qubit gates, 7.9 × 10−4 for two-qubit gates and 4.8 × 10−4 for state preparation and measurement. Richard W. Hamming, "Error Detecting and Error Correcting Codes", Bell System Technical Journal, volume 29, April 1950, for the opening. The four earlier Quentir posts linked inline, on the Bharti, Haug and Tanggara overhead theorem, on Princeton's quantum institute, on Roeland Wiersema's classical simulation of D-Wave's advantage experiment, and the Medicine Monitor's post on when the error check happens, are this site's own prior readings and carry their own source lists. The reading that a logical error rate is meaningful only beside the physical baseline, the qubit overhead and the protected operation set is Quentir's own.
Published intelligence, built to inform your own decisions. Published: September 4, 2026.