What record did Quantinuum break in error correction, and what does “10×” mean?

What did Quantinuum actually measure?

A logical memory experiment, not a computation. In the paper behind the announcement («Quantum error correction well below the color code threshold», arXiv:2610.24077, submitted October 16, 2026), 22 Quantinuum authors ran triangular color codes at distances 3, 5 and 7 on Helios-1, the company’s 98-qubit trapped-ion machine (what Helios is and what is measured). One logical qubit at distance 7 takes 73 physical qubits. The logical qubit idles while syndrome-extraction cycles run — 20 rounds per shot — and the failure rate per cycle falls as the code grows:

Each two-step increase in distance cut the logical error by a factor Λ = 12 (+3/−2) with the best decoder, Tesseract, run offline (arXiv:2610.24077). The press release translates the distance-7 point into «99.9995% logical fidelity» and «five errors per million operations» (Quantinuum, October 19, 2026). The unit behind that sentence is a syndrome-extraction cycle of an idling logical memory — not a logical gate, and not an application circuit.

«10×» better than what?

Better than Quantinuum’s own previous record, set the month before. In September 2026 the company published its Helix architecture — a [[20,2,6]] code holding two logical qubits in 20 ions — with a logical memory error of 4.6×10⁻⁵ per qubit per cycle on the same Helios hardware (arXiv:2609.03194, via Quantum Computing Report, September 9, 2026). The new distance-7 point sits at ≈5×10⁻⁶; the straight ratio is about 9, which the release rounds to «10×». Both numbers are real, both are Quantinuum’s, and both were measured on Quantinuum machines. Rosetta Quantum logs the claim with its baseline attached: this is a record against the holder’s own previous record — the second in two months, as the release itself notes — not against an external, neutral benchmark (how to tell if a claim is credible).

Is the «100× lower than any other modality» comparison like-for-like?

No. The release also claims the result is «100x lower than the best published result on any other hardware modality», linking Google’s surface-code paper. The linked figure is Google’s Willow processor holding a distance-7 surface-code memory at 0.143% error per cycle (Nature, December 2024). The straight ratio between 1.43×10⁻³ and 5×10⁻⁶ is about 286, so «100×» is, if anything, conservative — but the comparison crosses code families (color vs. surface), platforms (trapped ions vs. superconductors), machine sizes and decoder setups, and a per-cycle rate says nothing about wall-clock speed, where the two platforms differ sharply. Rosetta’s standing verdict applies: no two vendors publish logical-qubit demonstrations that can be compared like-for-like (what a logical-qubit demonstration really shows).

What sits between this record and fault-tolerant computation?

Three gaps, all documented in the paper itself. Gates: the record is for idling memory; the same paper’s logical two-qubit randomized benchmarking gives 7×10⁻⁵ per Clifford at distance 5 — 14 times the memory record — and could not be run at distance 7 because Helios-1 does not have enough qubits (arXiv:2610.24077). Decoding: every headline number was decoded offline, and the authors state that Tesseract is not yet fast enough for real-time use (why decoding is a real-time classical problem). Magic states: the paper demonstrates no magic-state preparation and no Clifford+T circuit — the ingredient that dominates the cost of useful fault tolerance (magic state distillation, explained). The release’s target — Apollo, a universal fault-tolerant machine, by 2029 — will remain a roadmap promise until those gaps close, and Rosetta will keep dating it as such (how late quantum roadmaps run).

Is the QLDPC surgery result the same record?

No — it is a separate preprint, on a different machine, and not authored by Quantinuum. On October 28, 2026, a two-author paper (Surabhi Luthra and Alex Owens, arXiv:2610.39067) reported individually addressed logical Pauli measurements by QLDPC surgery, run on Quantinuum’s H2-2 trapped-ion processor as the hardware platform. Secondary coverage credits the team to the UK’s National Quantum Computing Centre (Quantum Zeitgeist, November 1, 2026); the arXiv abstract page lists no affiliations. The demonstration used a [[10,2,3]] generalised bicycle code — 10 physical qubits, 2 logical qubits, distance 3 — plus 7 gadget qubits:

Why it matters: QLDPC codes are the main candidate route to fault tolerance with fewer physical qubits, and «surgery» — measuring logical operators without destroying the encoded block — is the primitive that turns them from storage into something you can compute with. This is a distance-3, percent-level validation of that primitive. It is nowhere near the memory record’s error rate, and it is not trying to be: it demonstrates an operation the color-code record does not attempt.

What will Rosetta watch next?

Rosetta will re-issue this verdict when one of three things happens: arXiv:2610.24077 passes peer review (both papers here are v1 preprints), a real-time decoding run of the distance-7 color code is published, or logical two-qubit benchmarking at distance 7 is measured on larger hardware. Quantinuum’s roadmap dates — Apollo by 2029 — will be tracked as promises, in future tense, until a machine shows up.

What we don’t know

• Which Google number the «100×» refers to: against the figure the release itself links (0.143% per cycle at distance 7, Nature, December 2024) the straight ratio would be ≈286×; the release does not show its arithmetic.
• How wide the record really is: three failures in 29,000 shots put the distance-7 point at 5 (+4/−2)×10⁻⁶ — a 1σ interval spanning almost a factor of 3 (arXiv:2610.24077).
• Whether Λ = 12 survives real-time decoding: every reported number was decoded offline (arXiv:2610.24077).
• Who the authors of arXiv:2610.39067 work for: the abstract page lists no affiliations; the NQCC credit comes from secondary coverage (Quantum Zeitgeist, November 1, 2026).
• Whether per-cycle memory comparisons across modalities predict anything about applications: no shared benchmark exists, and cycle times differ by platform.
• Whether either result will replicate outside Quantinuum hardware or authorship: no independent replication of the September or October results is published as of November 4, 2026.

Rosetta Quantum publishes verdicts with raw, reproducible data. This is educational content, not a product claim and not investment advice.