BRENNEN

Industries / Technology / Quantum computing

Error correction

Error correction spreads one unit of information, called a logical qubit, across many physical qubits and checks them constantly, so that errors are found and undone faster than they build up.

4

Milestones achieved, from 1996 to 2024.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1996

    The threshold theorem

    In 1995 Peter Shor published a scheme for protecting the memory of a quantum computer from decoherence, the loss of its quantum state. In 1996 and 1997 Dorit Aharonov and Michael Ben-Or, and separately Emanuel Knill, Raymond Laflamme and Wojciech Zurek, showed that if the error in each operation is below a fixed threshold, codes nested inside codes let a quantum computer run as long as needed and stay accurate. Experiments since, such as Google's below, test whether hardware is under such a threshold.

    api.crossref.orgarxiv.orgarxiv.org

    Credit

    Peter Shor (AT&T); Dorit Aharonov and Michael Ben-Or; Emanuel Knill, Raymond Laflamme and Wojciech Zurek

  2. 1997

    Codes laid on a surface

    Alexei Kitaev showed that a two-dimensional quantum system could protect information through its structure. In 2012 Austin Fowler, Matteo Mariantoni, John Martinis and Andrew Cleland set out how to build a full computer on the resulting surface code, in which qubits sit on a flat grid and each check involves only neighbouring qubits, and estimated its size and speed. The Google and USTC results below, and the experiments on rented chips under Open, use surface codes.

    arxiv.orgarxiv.org

    Credit

    Alexei Kitaev; Austin Fowler, Matteo Mariantoni, John Martinis and Andrew Cleland

  3. 2023

    48 logical qubits on atoms

    Dolev Bluvstein, Mikhail Lukin and colleagues ran a processor of up to 280 neutral atoms as logical qubits. Moving from a distance-3 to a distance-7 surface code (a code's distance measures how many errors it can catch; a larger distance needs more qubits) made a two-qubit logic gate better, and circuits on up to 48 logical qubits beat the same circuits run on bare physical qubits.

    Lessonthe 48-qubit circuits used codes that detect errors and discard the flagged runs rather than correcting them, a difference that later headline counts also turn on.

    arxiv.org

    Credit

    Dolev Bluvstein, Markus Greiner, Vladan Vuletić, Mikhail Lukin and colleagues

  4. 2024

    Below the threshold

    In 2023 Google reported that a distance-5 surface code slightly beat a distance-3 one, 2.914% against 3.028% errors per cycle. In August 2024, on its Willow chip, each step up from distance 3 to 5 to 7 cut the logical error rate by a factor of 2.14; the 101-qubit distance-7 code had 0.143% error per cycle, outlived its best physical qubit 2.4 times over, and was decoded in real time. In December 2025 USTC reported the same behaviour on its own 107-qubit Zuchongzhi 3.2 chip, a factor of 1.40 at distance 7, which Quantum Computing Report calls the first such result outside the United States.

    LessonGoogle found that rare bursts of errors hitting many qubits at once, about once an hour, then set the floor on how well its codes worked.

    arxiv.orgarxiv.orgquantumcomputingreport.com

    Credit

    Google Quantum AI; University of Science and Technology of China (USTC)

Watch for

  1. Five errors per million

    Quantinuum says a logical qubit on its Helios computer, built with a color code (another family of codes), reached 99.9995% fidelity, about five errors per million operations, ten times better than its own record of the month before. It says each step up in code distance cut errors by a factor of 12. Robin Blume-Kohout of Sandia National Laboratories, who was not involved, said the result shows "a logical qubit nearly 100x better than its component physical qubits".

    Achieved whenan outside group runs the same codes on Helios and gets matching error rates, or another group reaches a similar rate on its own hardware.

    hpcwire.com

    Claimed by

    Quantinuum, October 2026

  2. Magic states made cheaply

    A fault-tolerant computer needs a supply of special "magic states" to run all the operations an algorithm needs, and by some estimates up to 95% of the machine could be busy making them. Google reports in Science that magic state cultivation, a filtering method its researchers proposed in 2024, cut errors in these states about 40 times on Willow and gave states with 99.99% fidelity, then moved them from a small color code into a distance-5 surface code. Google says the method still has to be shown at larger scale.

    Achieved whenanother group runs cultivation on its own hardware and reports a similar cut in errors.

    blog.google

    Claimed by

    Google Quantum AI, October 2026

Open

  1. 01

    Logical-qubit claims, scored

    A public table that scores every logical-qubit claim since 2023 against the same five tests: whether the logical qubit beats its physical qubits, whether errors fall as the code grows, whether it holds over many correction cycles, whether it avoids post-selection (throwing away runs where an error was flagged), and whether it stays stable over useful lengths of time. Where data are public, recompute the headline number.

    The brief
    Credit · unclaimed
  2. 02

    Rented chip below threshold

    Run surface-code memory experiments at distances 3, 5 and 7 on a rented cloud processor with a square grid of qubits, such as IBM's 120-qubit Nighthawk, and show errors falling as the code grows in both bases (the two kinds of error a code tracks), without post-selection. Publish circuits, raw shots and decoder settings.

    The brief
    Credit · unclaimed

Open 01 · the brief

Logical-qubit claims, scored

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public table that scores every logical-qubit claim since 2023 against the same five tests: whether the logical qubit beats its physical qubits, whether errors fall as the code grows, whether it holds over many correction cycles, whether it avoids post-selection (throwing away runs where an error was flagged), and whether it stays stable over useful lengths of time. Where data are public, recompute the headline number.

Why it is still open

In June 2026 Alice & Bob, a company that builds cat qubits, published these five criteria as a proposed standard. Claims mean different things: PostQuantum reports that Infleqtion's 30 logical qubits of September 2026 rely on error detection, in a distance-2 code with post-selection, which finds errors but cannot correct them. A November 2026 search found no table that applies the criteria to past claims.

thequantuminsider.compostquantum.com

Built with what exists

The papers and their public datasets, such as Google's; Stim, Google's open-source library for simulating error-correction circuits; PyMatching and the open decoders released with the 2026 "Sim-to-Real" study; a spreadsheet.

github.comarxiv.org

  1. First step

    Score five claims: Google 2024, USTC 2025, Quantinuum September 2026, Quantinuum October 2026 and Infleqtion September 2026.

  2. Done looks like

    At least 20 claims scored on all five tests, each score linked to the page that supports it; for at least 5 claims with public data, the logical error rate recomputed independently; each company shown its row before publication and its reply recorded.

Open 02 · the brief

Rented chip below threshold

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

Run surface-code memory experiments at distances 3, 5 and 7 on a rented cloud processor with a square grid of qubits, such as IBM's 120-qubit Nighthawk, and show errors falling as the code grows in both bases (the two kinds of error a code tracks), without post-selection. Publish circuits, raw shots and decoder settings.

Why it is still open

Below-threshold results so far come from the machines' builders. Outside teams ran distance-5 surface codes on 120-qubit square-lattice processors in 2026. Julien Drouet, Robin Harper and colleagues cut errors to 1.62% per round by leaving out faulty qubits and couplers, but their distance-5 code beat the best distance-3 code in only one basis, and only with post-selection. Moritz Lange, Mats Granath and colleagues ran a neural decoder on an IBM Nighthawk-generation processor; their paper reports gains over a standard decoder, not errors falling as the code grows.

arxiv.orgarxiv.org

Built with what exists

IBM Quantum Platform (pay-as-you-go from USD 96 a minute), Qiskit, Stim, PyMatching, and the defect-exclusion methods of Drouet and colleagues.

ibm.com

  1. First step

    Rerun the Drouet and colleagues distance-5 memory experiment on a current Nighthawk r2 and compare with their 1.62% per round.

  2. Done looks like

    The logical error per round falls from distance 3 to distance 5 in both bases, with error bars, without post-selection, from data that a second team decodes independently and gets the same result.