QCD Original

Three quantum records, three different rulers: what 99.76%, 0.72% and 256 qubits actually count

What a Princeton gate fidelity, a trapped-ion logical measurement error and a 256-qubit software proof each measured, why none predicts algorithm performance, and a checklist for decoding benchmarks.

A Quantum Computing Digest Original StoryThis article was researched and written entirely by AI, without human review or editing, as part of the Quantum Computing Digest's ongoing experiment in AI-powered journalism.

Three quantum computing results from late October 2026 measured three different things, so their headline numbers cannot be put on one scale. Princeton's 99.76% is the success rate of a single operation between two atoms. The trapped-ion experiment reports the error of one measurement on encoded data. The 256-qubit 'verification' is a mathematical check of circuit designs that never ran on a quantum computer.

None of the three tells you how well a machine would run a full algorithm. All three are arXiv preprints that have not yet been peer reviewed. Below we explain what each number counts and give a short checklist for reading the next benchmark headline.

A gate measured two qubits at a time

The Princeton paper, posted on 29 October by a team that includes Jeff D. Thompson and Waseem S. Bakr, describes an array of caesium-133 atoms kept in a cryogenic (deeply cooled) environment. The authors report a single-qubit gate fidelity of 99.970% and a 'raw' two-qubit entangling-gate fidelity of 99.76%, with an uncertainty of 0.08 percentage points. Fidelity means how closely an operation's actual result matches the intended one, averaged over many repetitions. A two-qubit gate is the basic operation that links two qubits so that their states become correlated, or entangled.

One news headline called this '99.76% quantum entanglement'. What the paper reports is a gate fidelity, which describes one operation on a pair of atoms. The reason for cooling is physical. The gates rely on Rydberg states, highly excited atomic states that let neighboring atoms interact, and blackbody radiation (the heat glow from warm surroundings) disturbs those states. The authors say a cryogenic environment can cut those disturbances by more than a factor of ten. They put the error contributed by Rydberg decay below 0.03% and say most of the remaining error comes from atomic motion, which they expect better cooling to reduce.

The figure does not tell you how many gates can run in a row, how many atoms work at once, or how errors behave across a large array. The Quantum Zeitgeist write-up itself warns that raw two-qubit figures leave out some error sources that appear in larger systems. Simple arithmetic shows the gap. An error of 0.24% per gate means that, if errors were independent, a circuit of 1,000 such gates would finish without a single gate error only about 9% of the time.

One measurement, not one qubit

The researchers used Quantinuum's H2-2 trapped-ion processor, which has 56 physical qubits. They stored two logical qubits in a block of 10 physical data qubits. A logical qubit is one unit of information spread across many physical qubits so that errors can be detected and fixed. The block used a quantum low-density parity check (QLDPC) code, a family of codes that packs several logical qubits into one block. The experiment used 27 physical qubits in total.

The test was 'QLDPC surgery', a generalization of lattice surgery. Extra 'gadget' qubits are briefly attached to the code block so that one logical qubit can be measured without disturbing the other. A single round of this measurement had an error rate of 0.72%. The 95% credible interval, the range the true value very likely falls in, runs from 0.32% to 1.41%. Repeating the check over three rounds gave zero errors in 957 runs, which puts the error below 0.26%. The authors note two caveats. The single-round version has a circuit distance of 1, meaning one fault can cause an undetected logical error. They also discarded runs that failed a state-preparation check, fewer than 5% of shots.

So the reported figure is the error of one logical measurement, not the error of a logical qubit over time. The authors' own error model for the device gives a 0.1% physical two-qubit gate error. The experiment was meant to show that the measurement step works, not to beat physical qubits. The authors say decoding was done offline, after the runs. They list real-time decoding, multi-qubit measurements and larger codes as next steps. They argue that QLDPC codes could reduce the overhead of surface-code designs, which they say need hundreds to thousands of physical qubits per logical qubit.

256 qubits that never ran

The third headline, 'Researchers Verify Quantum Simulation For 256 Qubits', covers a computer science paper filed under programming languages. Its authors are Min Lai and Shih-Han Hung of National Taiwan University and Yu-Fang Chen of Academia Sinica. The paper presents software for formal verification, meaning a mathematical proof that a program matches its specification. It applies this to circuits for Hamiltonian simulation, the task of computing how a quantum system changes over time. The tool checks circuit designs written for up to 256 qubits, with an estimated size of about 1.52 million gates, without building the huge matrix that would describe them directly.

No quantum hardware ran, so there was no machine output to verify. The 256 is the width of circuits that the software proved correct by design; it is not a count of physical or logical qubits that ran anything. Proofs like this can catch bugs in quantum programs before they run, but they say nothing about how accurately a real machine would carry them out.

A checklist for the next headline

What was measured? It could be a single physical gate, a logical operation, a logical qubit stored over repeated cycles, or a piece of software.

How many qubits, and of which kind? In these three cases the answers were two physical atoms, 27 physical qubits holding two logical qubits, and 256 qubits on paper.

Per what? A percentage means little without its unit, whether that is per gate, per measurement, per round of error correction or per complete run. Also check whether results were raw or filtered: look for discarded runs, error bars and the number of repetitions. A few hundred or a few thousand runs leave wide uncertainty.

Did anything run end to end, and has the work been peer reviewed? For all three results here, the answers are no and not yet. Together the papers show better physical gates, a working logical operation and better tools for checking programs. But the public record has no formula for turning one of these numbers into another. A fair comparison would need the same quantity, measured the same way, on the same task.