WorksPhysics and Astronomy
Resolving and Suppressing Temporally Correlated Noise in Quantum Error Correction
Abstract
Scaling fault-tolerant quantum error correction (QEC) requires suppressing errors in both space and time, yet standard logical memory experiments primarily probe spatial faults. Here, we use a 105-qubit Willow superconducting quantum processor to realize and scale the surface code stability experiment, which is alternatively sensitive to temporal error chains. Scaling from patch sizes of to , we realize a timelike block suppression factor of , demonstrating the ability to maintain temporal error suppression as the number of qubits increases. By extending the stability experiment for many cycles, we uncover a sum-of-exponentials behavior in error suppression caused by a temporally correlated error: data-qubit heating to the third excited state . We fully capture this effect with a parameter-free, device-level leakage model incorporating higher-state transport during syndrome extraction cycles. To suppress this error, we develop and realize a walking stability circuit, reaching a logical error probability floor of at 60 cycles—two orders of magnitude below the standard construction—and limited by ionizing radiation error bursts. Finally, by adaptively terminating syndrome extraction based on decoder confidence in an offline analysis, we reduce spacetime logical error by 38% and time overhead by 34% compared to the optimal fixed-cycle strategy.
Introduction
Quantum error correction (QEC) will allow the realization of fault-tolerant quantum algorithms using noisy physical qubits by suppressing logical error rates far below physical error rates [6]. The standard experimental benchmark in QEC is the so-called memory experiment, in which a logical qubit is initialized, repeated cycles of error correction are performed, and a decoder is used to correct the logical observable measured at the end of the experiment [4, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17]. However, fault-tolerant computation will involve not only qubit preservation, but also controllable logical operations on encoded qubits. Recent experiments have focused on realizing and modeling these fault-tolerant logical operations [18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31], which can often have vastly different sensitivities to physical error mechanisms relative to memory experiments. It is thus crucial to benchmark, model, and suppress all physical error mechanisms that will limit fault-tolerant operations, as a focus on memory alone can lead to an over-optimization of system design toward logical qubit preservation and is blind to architectural tradeoffs that may improve memory while worsening the fidelity of logical operations.
Notably, memory experiments are often most sensitive to spatial physical errors that occur across multiple qubits at a similar time—the worst being high-weight errors spanning multiple qubits and flipping an observable. Logical operations, on the other hand, can have regions in spacetime that are instead most sensitive to temporal errors that occur at a similar spatial location across many cycles of error correction. Benchmarking device performance using simple experiments that probe logical sensitivity to both spatial and temporal errors would allow for a window into both of these mechanisms. Importantly, uncovering any low-probability correlated errors now (an example being high energy events [4, 8, 32, 33]) will allow for hardware mitigation before scaling up.
The stability experiment [5] is the spacetime dual to the memory experiment, and is most useful for benchmarking the impact of temporal errors on logical operators. Importantly, it can be used together with memory experiments to predict the performance of primitive logical operations such as lattice surgery and logical qubit moves [34]. As such, the stability experiment is another crucial benchmark that should be run on all devices and should be treated on equal footing with the memory experiment for scaling up logical operations. Recent experiments have begun exploring stability and its repetition-code-like variants [14, 15, 16, 28, 30].
In this work, we scale surface code stability experiments up to a full-device patch and hundreds of error-correction cycles on Google’s 105-qubit Willow processor. Using the stability experiment alongside memory experiments, we introduce and benchmark the spacetime error, which is sensitive to both temporal and spatial errors. Using device-informed models, we reveal the relative importance of various physical error channels between spacelike and timelike observables, including the relatively high importance of readout and reset in the stability experiment. Additionally, we use the stability experiment as a metrology tool, uncovering low-probability temporally correlated errors. We then fix these errors in experiment using a walking stability circuit [10, 35], realizing a minimum logical error probability of at 60 cycles—suppressing logical error by two orders of magnitude compared to the standard circuit compilation at 60 cycles. Finally, we use the decoder to analyze dynamic stopping strategies ex post facto. Using soft information from the decoder [36, 37], we demonstrate a 38% reduction in error and a 34% reduction in mean runtime when minimizing error.
Stability experiment regimes
To understand the stability experiment, we first compare it to the more well-known memory experiment. A standard patch of surface code encoding a distance logical qubit for a memory experiment is shown on the left of Figure 1a. In this patch, data qubits and measure qubits are aligned in a square grid, with the top and bottom (left and right) boundary weight-2 stabilizers being Z-type (X-type). In such a configuration, the logical operators () are defined as products of data-qubit X (Z) operators connecting the left and right (top and bottom) edges of the patch [2]. In contrast to the memory experiment, the stability experiment’s patch of surface code does not encode a logical qubit, as shown in the right two grids in Figure 1a. Instead, stability has only one type of boundary stabilizer defining the basis of the experiment. In such a configuration, the -basis stabilizers () are overdetermined, and their product satisfies the global constraint , defining a logical observable . The stabilizers of these patches are repeatedly measured for cycles, and the results are fed to a decoder to correct errors in the measurement of , which can be defined at any arbitrary cycle. Repeatedly measuring the stabilizers for cycles defines a decoding graph for which errors across time can be corrected, and hence the stability experiment has a fault distance .
Figure 1. a) Memory and stability surface code patches. Measure (data) qubits are small (large) dots. X-type (Z-type) stabilizers are colored red (blue). The X patch has cut corners in order to maintain the checkerboard in the bulk, which sets the gate ordering. b) Measured stability logical error probability on 105-qubit Willow with correlated matching decoding. Data are averaged across 7, 4, 2, 1, and 1 subgrids for patch width and , respectively, with each subgrid and basis running a 200-cycle stability circuit for raw shots. Extracting five subsampling windows per shot yields () effective shots per circuit for (). *Due to device constraints, and grids are Z-basis stability experiments only; all other grids are averaged over Z and X bases. Inset: quadratic scaling of logical error with patch width after 5 cycles. c) Illustration of Pauli and correlated error regimes.
To benchmark the performance of stability in hardware, we run stability experiments with data-qubit width height from to on a 105-qubit superconducting Willow processor [4]. Similar to memory scaling experiments, we average the logical results of multiple stability patches of each size across the device to accurately compare the results of low-patch-width experiments (e.g., ) to full-device experiments (). We use the standard 4-layer stabilizer measurement circuit compiled to CZ operations and idle between cycles in the ZXXZ code to remove unnecessary single-qubit gates [4]. We also include dynamical decoupling, measure-qubit leakage removal using multi-level reset (MLR) [38], and data-qubit leakage removal (DQLR) of the state using leakage-iSWAP and fast measure-qubit reset operations [9].
The measured results of our stability-scaling experiment across patch sizes, decoded with a correlated matching decoder [39] and an SI1000 error model [40] prior, are shown in Figure 1b, where the logical error probability is averaged over multiple patches of the same size and across bases (Z and X) [41] (see SM secs. S1 and S10). To increase the effective number of shots, we employ a subsampling technique in which short-duration stability-experiment data is taken as multiple windows from longer stability experiments (200 cycles for the scaling dataset and 1000 cycles for the walking experiment) (see SM sec. S8). We calibrate our experiment using reinforcement learning [42] with a cost function that includes both memory and stability detection probabilities (see SM sec. S9).
Two prominent features are shown in the data. First, unlike in a memory experiment, the measured stability logical error probability at a fixed number of cycles worsens as patch size increases in space, as shown in the inset of Figure 1b. This is expected from a simple path-counting argument; the number of possible error chains that can cause a logical error increases approximately linearly with the number of qubits, and the resulting logical error probability at a fixed number of cycles increases quadratically with respect to the patch width.
Second, our data differs from a single exponential suppression with cycles, which is expected for uncorrelated local errors given the distance increase with cycles. After around 10 cycles, the logarithm of the logical error probability, (LEP), incurs a knee, after which the suppression worsens. To quantify performance in both regimes, we fit our data to a sum-of-exponentials model, , as motivated under a simple correlated error model (see SM sec. S2), with results given in Table I. Notably, the exponential suppression after the knee, , is consistent across patch widths, indicating the same mechanism is likely limiting suppression after the knee for all patches. As will be shown through both detailed noise modeling and direct measurements in Section IV, this sum-of-exponentials behavior is caused by data-qubit leakage to the third excited state , which dominates logical errors after the knee.
| Patch Size | () | |||
| 3 3 | ||||
| 4 4 | ||||
| 5 5 | ||||
| 6 6 | ||||
| 7 7 |
Table I. Fit parameters for the two-component stability decay model across patch sizes. Uncertainties represent standard errors.
The stability experiment and its sum-of-exponentials behavior can be understood in terms of metrology with two regimes. The Pauli regime (before the knee) is defined as the section where logical error can be accurately predicted with a stochastic Pauli model. The correlated error regime (after the knee) is dominated by low-probability non-Markovian error mechanisms. Potential sources of these errors include higher-state leakage, measurement-induced state transitions, and coupling to coherent two-level system (TLS) defects. Under some correlated error models, the logical error can exhibit multiple knees, with regions where the logical error worsens with cycles before being further suppressed (see SM sec. S2 A).
Behavior in the Pauli regime
Near-term experiments involving logical gates will require a detailed understanding of both the spacelike and timelike error mechanisms in the regime where cycles . Hence, we benchmark Willow in this regime using both stability and memory experiments (see Figure 2a) on multiple patch sizes tiling the device. Here, correlated errors are subdominant, as shown by our stability data in Figure 1b; as a result, a stochastic Pauli model accurately predicts device performance in this so-called Pauli regime. We show this by constructing a circuit-level stochastic Pauli noise model directly from per-qubit and per-pair device benchmarks [10] (see SM sec. S4). Coherent errors and single-cycle leakage effects are Pauli-twirled into effective depolarizing and bit-flip channels before entering the model. Simulating this Pauli model across both and bases reproduces the measured logical error probabilities for both memory (–) and stability (–) at cycle counts comparable to the code distance (see Figure 2b), including the characteristic even-odd cycle staircase in short stability experiments. However, the stability data diverges sharply upward from the Pauli prediction after cycles for patch widths , respectively. This knee occurs when the cumulative probability of a multi-cycle correlated error event exceeds the exponentially suppressed stochastic Pauli logical error probability.
Figure 2. a) Memory and stability experiment topological diagrams. The colored red and blue sheets represent and logical observables. b) Measured memory and stability experiments in the Pauli regime, compared to a Pauli model simulation. c) Measured and Pauli-modeled spacetime error probabilities, with minima at the optimal cycle count shown as stars. d) Scaling of the block logical error probabilities (, , and ) evaluated on a spacetime block with patch width ( for memory). e) Inverse block error suppression () budgets for memory, stability, and spacetime under a Pauli model. f) Comparison of the relative contribution of various error channels between stability and memory, along with the ratio between stability and memory logical error probabilities at the optimal cycle count ( at ).
To quantify the performance relevant for logical operations such as lattice surgery and logical qubit moves, we define the spacetime error probability . For both memory () and stability () on the same patch of qubits, we take the sum of the logical error probabilities across the and bases, and ; the spacetime error is then defined as the average of these sums, . For our data, is plotted in Figure 2c. Since the stability error is exponentially suppressed with cycles, and the memory error increases linearly with cycles to first order, a minimum in occurs at an optimal cycle count . To avoid statistical uncertainty when identifying from single-cycle data points, we determine by finding the minimum of the fitted spacetime curve obtained from the logical error rate fit to and the two-exponential decay fit to . We call the resulting minimum the minimum spacetime error , where and denote the constituent memory and stability block error probabilities evaluated at the optimum . We find the optimal number of cycles for patch widths to be , indicating that logical operations such as lattice surgery and qubit moves will likely benefit from waiting or cycles, which is longer than the typical rule of thumb , and consistent with the balance between timelike and spacelike error mechanisms presented in [34].
Under stochastic Pauli noise, the minimum spacetime error is expected to be exponentially suppressed with increasing patch distance (). As such, we fit the minimum spacetime error using the suppression model , and find a spacetime suppression factor of , indicating Willow can suppress the error rate of logical operations that are sensitive to both spatial and temporal errors as the number of qubits is increased. Only odd-distance points are included in the fit, and the symbol denotes a block suppression factor rather than a per-cycle suppression factor. We expect better performance can be achieved using the same data under more sophisticated decoding techniques, including ML decoding [43].
The measured spacetime suppression is slightly lower than the Pauli-model prediction of (Pauli model) . This difference is due to error mechanisms such as leakage, which are not captured by the Pauli model. Accordingly, we introduce a full leakage model in the next section. In Figure 2d, we compare the spacetime error suppression to the spacelike block suppression ( (Pauli model) ) and timelike block suppression ( (Pauli model) ). All block error rates in Figure 2d (, , and ) are evaluated at the same optimal cycle count , where a logical block is defined by a spacetime volume of , with for the memory experiments.
Finally, to quantify how individual physical error mechanisms limit performance in the Pauli regime, we construct error budgets for the spatial (), temporal (), and combined spacetime () inverse block error suppression factors. All such error budgets are evaluated at the fixed experimental optimal cycle counts via one-at-a-time channel ablation in our Pauli model (Figure 2e). By ablating each error channel family in turn, we partition each into contributions from two-qubit gates (2Q Gates), data-qubit idling during measurement and reset (Idle), mid-circuit readout, reset, single-qubit gates (1Q Gates), and data-qubit idle errors during leakage removal operations (DQLR). While two-qubit gates (55% in memory, 43% in stability) and data-qubit idling (32% and 25%) represent the largest absolute contributions to both metrics, their relative contributions differ markedly (Figure 2f). Temporal chains of measure-qubit flips directly penetrate the stability observable sheet ; hence, predominantly timelike error channels carry increased fractional weight in stability when compared to memory—specifically reset (6.0% vs. 2.5%, a 2.43× relative contribution) and readout (21.0% vs. 4.9%, 4.28×). Conversely, predominantly spacelike data-qubit and other error channels carry reduced or roughly equal relative weight in stability (‘2Q Gates’ at 0.78×, ‘Idle’ at 0.77×, ‘DQLR’ at 0.53×, and single-qubit gates at 1.05×).
At the minimum spacetime error , the optimal cycle count occurs where the steep exponential decay of stability balances the gentler linear growth of memory with cycles; as a result, the stability error is substantially smaller than the memory error at the optimum, as seen directly in the block error rates of Figure 2d. Specifically, stability () accounts for 16% of the combined logical error across odd distances compared to 84% from memory (, corresponding to a ratio in Figure 2f); consequently, in the Pauli regime, the combined spacetime error budget () closely mirrors the memory budget (54% two-qubit gates, 31% idling, 6% readout, 4% single-qubit gates, 3% reset, and 2% DQLR), while remaining first-order sensitive to any degradation in temporal error suppression. The associated budgets and single-channel impact on logical error probability for memory, stability, and spacetime are detailed in SM sec. S5.
Behavior and modeling in the correlated error regime
Although near-term logical performance on smaller surface code patches () is governed by Pauli errors, time-correlated errors will limit performance at larger cycle counts, and thus larger logical patch sizes. In order to build intuition for how such error mechanisms impact the performance of the stability experiment, we consider a minimal phenomenological model in which a local defect occurs with probability per qubit per cycle and persists for a characteristic lifetime cycles. The defect, while active, will randomize adjacent stabilizer measurements (see SM sec. S2); a realization that persists for cycles is sufficient to cause a logical fault. This is observed as the stability logical error probability transitions from steep Pauli suppression to a shallow correlated regime governed by the tail of the defect lifetime distribution. In this two-regime sum model, the logical error at which the knee appears scales quadratically with patch width and linearly with the defect occurrence rate (), whereas the post-knee slope is set by the persistence lifetime , yielding a severely degraded post-knee suppression factor ( across – in Figure 1b). Notably, if the defect duration is concentrated near a fixed length , the stability logical error probability can increase non-monotonically with cycle count for before decaying.
While this simple model suffices for capturing the high-level intuition, mitigating such errors on the physical device requires elucidating their underlying physical mechanism. Although Data Qubit Leakage Removal (DQLR) implemented on Willow processors [9] resets the population every cycle with high efficacy, it is not explicitly designed to remove the population (see SM sec. S3). Consistent with this, we show that the dominant time-correlated error is the low residual population, produced primarily by passive heating within the device. Heating rates were characterized on the device and included directly within our model; a detailed description and measured rates can be found in SM sec. S3. Phenomenologically, leakage to higher excited states behaves similarly to the defect described above. Once in , a data qubit will corrupt adjacent stabilizer measurements across consecutive cycles until it exits the manifold. We note that aside from intrinsic relaxation, the dominant removal mechanism is leakage transport during diabatic CZ gates, in which leakage swaps onto the measure qubit, where it is cleared by multi-level reset (see SM sec. S3).
Incorporating all measured gate component benchmarks and heating rates alongside analytical leakage transport into a single model reproduces the measured stability logical error probabilities across all patch widths (–) and out to 70 cycles without any free fitting parameters (Figure 3a). Beyond aggregate logical performance, the multi-level model reproduces both the ensemble distribution and single-shot spatiotemporal structure of syndrome detections (Figure 3b,Figure 3c).
Figure 3. Stability in the Correlated Error Regime. a) Leakage model prediction (lines) for stability logical error probability vs. cycles, compared to experimental data (triangles). b) Cumulative distribution of per-detector firing probabilities across a () stability experiment (solid teal, median 0.0726) and the multi-level simulation (dashed black, median 0.0756). c) Representative individual shots of detection patterns from experiment (left) and simulation (right), showing multi-cycle syndrome bursts across the four measure qubits bordering a -leaked data qubit (shaded yellow band). d) Final data-qubit state after 14 cycles in the stability experiment, postselected on limiting correlated error chains.
Across all detectors in the patch, the simulated detection probabilities match the measured distribution with a median of 7.56% compared to 7.26% in experiment (Figure 3b). At the single-shot level, the simulated realizations exhibit the same qualitative signature as experimental shots near the error floor: isolated background detections punctuated by abrupt, multi-cycle bursts of high detection density across up to four measure qubits adjacent to a single -leaked data qubit (Figure 3c). These correlated syndrome patterns persist for – cycles—consistent with the effective residence time set by the action of intrinsic relaxation and leakage transport.
For primitive logical operations such as lattice surgery and qubit movement requiring cycles on a patch of width , experiments at operate in or near the Pauli regime (), where stability () contributes only 16% of the minimum spacetime error . However, as shown by the model, once the code distance increases past , a logical operation of duration crosses the stability knee into the -dominated correlated error regime. We note that despite spacelike memory errors decreasing rapidly with distance (the memory logical error rate follows with suppression factor in the Pauli model and 1.98 in experiment), we expect post-knee stability errors to grow quadratically with patch width () and decay slowly with cycle count. In short, the correlated timelike error is expected to overtake the memory error at – (–), pinning the asymptotic spacetime suppression factor to unless leakage is actively mitigated.
Lastly, we confirm the model’s prediction that the state is limiting the logical error probability in the correlated error regime through direct measurement. Specifically, we run a short ( cycles) stability experiment on a surface code patch and measure data qubits at the terminal layer of the circuit with a readout that resolves between the , , , states. We investigate the final data-qubit state using data-qubit IQ values postselected on two conditions used to isolate temporal error bursts; specifically, we postselect on (i) logical failures and (ii) shots for which the detection fraction in the second half of the shot is higher than in the first half. The latter condition allows us to nominally catch shots when the limiting correlated error chain runs into the end of the experiment.
The results of our direct data-qubit readout investigation are shown in Figure 3d. Notably, out of flagged events, a majority (52%) terminate with the flagged data qubit in the third excited transmon state , whereas none terminate in , using simple readout classification. This stark asymmetry between and arises from our use of -state DQLR in the experiment, without any engineered mechanism to remove the state. We also observe an outlier measurement, which may indicate the presence of even higher () data-qubit states and is the subject of ongoing investigation. Some leftover population in and is expected from our imperfect shot-picking heuristic; however, the strong presence of the data-qubit state in shots with correlated error chains further confirms it as the limiting error mechanism in the correlated regime on this device, in addition to the agreement with the leakage model.
Mitigating temporally correlated errors via walking
The walking syndrome extraction circuit was introduced in [35] and experimentally realized for a memory experiment in [10]. The walking circuit is a dynamic QEC protocol that alternates between two distinct cycle types. This alternation enables a key property: by shuffling logical information between data and measure qubits, the circuit reads out and periodically resets all qubits. As a result, when using a multi-level-reset scheme that resets higher qubit states (including ), leakage is removed from all qubits periodically. Given that the state is limiting in the correlated error regime, a walking circuit can mitigate this floor by periodically removing the state on all qubits when a multi-level reset is used. In this work, we design a QEC circuit that realizes the walking stability code by modifying the boundaries of the memory walking circuit given in [35].
We experimentally run the walking stability circuit, with results compared against the standard (non-walking) stability circuit of the same size, as shown in Figure 4a. There is an expected overhead associated with walking given the larger qubit count, leading to a higher logical error probability in the Pauli regime. Here, the walking circuit lowers the suppression rate, as expected from a Pauli model simulation (see SM sec. S7), yielding a logical error probability roughly an order of magnitude higher at 10 cycles. Although we incur this overhead, the walking circuit successfully mitigates the knee from the state, reaching a logical error probability upper bound of at 60 cycles—a two-order-of-magnitude improvement over the non-walking circuit. Specifically, at cycles, the walking stability circuit exhibits 1 logical failure out of non-overlapping subsampled shots ( logical error probability), run over 5.55 hours of circuit execution time.
Figure 4. a) Measured walking (blue) vs. standard non-walking (red) stability logical error on a patch of surface code. Walking stability data run a 999-cycle Z-basis circuit on a single subgrid for raw shots. Extracting non-overlapping subsampling windows yields effective shots for (one window per shot) and () effective shots at () from 27 (8) tiled windows per shot, with error bars spanning binomial probabilities within a factor of 1,000 of the maximum likelihood (, showing upper bounds where zero errors occur). b) Windowed detection fraction for the error burst causing the walking floor. The red region highlights a failed subsample.
We find that, like the repetition code, the walking stability circuit is limited by ionizing radiation bursts [4, 8, 32, 33, 44]. As shown in Figure 4a, the walking stability floor is caused by a single error burst, with detection statistics for this burst shown in Figure 4b. It is likely that this second knee can be surpassed by extending the stability experiment beyond cycles, since bursts typically last on a timescale of ms.
Adaptive stopping with decoder confidence
Unlike memory experiments, stability experiments allow protection against timelike errors to be chosen adaptively by deciding when to terminate [37]. In a fault-tolerant quantum computer, terminating early after clean cycles of syndrome extraction can accelerate logic, while delaying termination after noisy cycles can avoid time-like logical failures and substantially reduce the logical error rate. To make this decision quantitatively, we use a measure of decoder confidence called the complementary gap: the log-likelihood ratio of the two most likely sets of errors giving different observable outcomes as estimated by the decoder [36]. A gap of zero means the decoder is entirely unsure of its prediction—effectively a coin flip—while a large gap indicates high confidence. Using an offline analysis, we compute the gap after each cycle, and terminate once it exceeds a syndrome-dependent threshold (see SM sec. S6).
In a representative stability shot (Figure 5a), we observe a sustained burst of detection events neighboring a single data qubit between cycles 6 and 13. This burst confuses the decoder, keeping the complementary gap small. Once the burst subsides, the decoder changes its prediction and, by cycle 20, has become confident in the new one. With the resulting flagged threshold of 21 nats, we wait out the burst and terminate with the correct prediction. Adaptively deciding when to stop reduces the mean stopping time by and the spacetime logical error rate by . It has a considerably greater effect on the timelike logical error probability—a reduction given a mean hold time of 12 cycles—but this offers diminishing returns when coupled with accumulating spacelike error at long hold times. Still, this dramatic reduction in timelike error suggests more significant spacetime logical error improvement as we move to larger distances in the presence of long temporally correlated error sources.
Figure 5. a) An example stability experiment. Activated detectors are highlighted. Measure qubits adjacent to a candidate leaked data qubit are highlighted in red. b) The random detection event pattern confuses the decoder, keeping its confidence low. After the pattern clears, the decoder gains confidence, eventually switching its decision (sign) and committing to it. c) Comparing fixed to adaptive strategies in the standard surface code and the walking surface code for a patch. d) Estimating the spacetime logical error probability on the grid by combining the timelike stability outcome of each shot at its stopping cycle with the accumulated spacelike memory error over cycles. With adaptive cycles, 50% of shots complete in cycles, 90% complete in , and 99% in .
The improvement in performance does not come for free. First, adaptively deciding when to stop round-to-round requires fast decoder decisions and feedback on the order of . This is conceivable but would be substantially alleviated by making decisions only every several cycles or using a simplified heuristic, both of which will affect performance. Also, while the mean hold time is reduced, there can be long hold times lying in the tail of the hold-time distribution. Practically, compiling an algorithm with variable-length gates that can take quite long will be challenging. Nevertheless, engineering the ability to wait out rare bad events may prove important for building a large-scale superconducting processor.
Conclusion
In this work, we realized stability and memory experiments on a 105-qubit Willow processor to benchmark both spatial and temporal errors in quantum error correction. By combining memory and stability logical error probabilities into the spacetime error metric, we measure a spacetime suppression factor of as the patch size scales from to in the low-cycle Pauli regime. Moving forward, device performance improvements will improve this suppression factor, and we estimate that using IQ data (soft information) in decoding, along with a more sophisticated decoder, will improve the stability performance [43, 45].
Extending the stability experiment to higher cycle counts reveals a sharp knee in logical error probability caused by data-qubit heating to the third excited state . Because our data-qubit leakage removal (DQLR) operation is calibrated to remove -state leakage via a leakage-iSWAP interaction without targeting the state, data qubits that heat to remain in for roughly ten cycles until cleared by intrinsic relaxation or leakage transport to measure qubits during CZ gates, reducing the measured Pauli-regime temporal suppression factor from the Pauli-model prediction of to 1.47. A device-level multi-level simulation with no free parameters reproduces this knee across all patch widths and shows that, if left unmitigated, -state leakage will overtake spatial memory errors at code distances –13. Finally, by implementing a walking stability circuit in which every qubit alternates between data and measure roles and undergoes a multi-level reset (which resets higher excited states including ) every two cycles, we remove the -state knee and reach a stability logical error probability floor of at 60 cycles, limited by ionizing radiation error bursts and suppressing logical error by two orders of magnitude compared to the standard compilation at 60 cycles.
While the walking circuit eliminates the -state knee, it introduces higher logical error rates at low cycle counts than the standard circuit (Figure 4a). Several alternative strategies could mitigate -state leakage without walking the surface code patch. First, DQLR could be extended to also transfer -state population from data qubits to measure qubits prior to measure-qubit reset, while leaving the computational subspace undisturbed. Second, reducing passive heating rates into during data-qubit idling and dynamical decoupling would directly lower the onset probability of the knee. Third, at the decoding level, monitoring the complementary gap after each cycle allows the duration of a stability experiment to be chosen adaptively—terminating early after clean cycles or running additional cycles to wait out temporal error chains—yielding considerable improvement in timelike logical error (Figure 5).
Together, these results show that memory experiments alone are not sufficient to predict the performance of fault-tolerant logical operations. Primitive operations such as lattice surgery and logical qubit moves require cycles of stabilizer measurement and contain spacetime regions sensitive to both spatial and temporal errors. In the Pauli regime, our spacetime error measurements indicate that logical operations benefit from waiting to cycles rather than , while our error budgets show that mid-circuit readout () and reset () carry substantially higher relative weight in stability compared to memory (Figure 2e,Figure 2f). More broadly, because the fault distance of a stability experiment is set by the number of cycles () rather than the number of physical qubits, even small surface code patches can probe temporal error floors down to , providing a practical metrology tool for identifying and eliminating correlated error mechanisms before scaling to larger fault-tolerant processors.
Note added.—While finalizing this manuscript, we became aware of [46], which realizes a similar circuit construction to § 1.4.
Contributions
A. Eickbusch led the project, performed the experiments, analyzed the data, developed the simple modeling, and realized the walking code. N. Noll led the modeling effort, developed the detailed modeling, built the leakage simulation, and analyzed the impact of leakage to higher excited states as a time-correlated error. A. Eickbusch and N. Noll conceived of the project. A. Eickbusch, N. Noll, and M. Newman wrote the paper with input from all the authors, and A. Morvan provided support on the manuscript. M. Newman created the gap figures and adaptive stopping analysis, wrote the corresponding supplement, and together with C. Gidney provided overall guidance throughout the project. M. McEwen guided the spacetime logical error analysis. L. Martin developed and implemented echoes in the walking code. G. Roberts, N. Lacroix, A. Bourassa, E. Rosenfeld, A. Bengtsson, K. Satzinger, and R. Cortiñas assisted in code infrastructure development and calibrations. V. Kurilovich, Y.-C. Wei, and G. Roberts assisted with the analysis of error bursts. C. Ding, L. De Lorenzo, and F. Borjans developed infrastructure and measured heating rates and population transport. V. Sivak developed the reinforcement learning calibration. A. Korotkov, S. Boixo, K. Miao, A. Opremcak, Z. Chen, C. Jones, J. Kelly, D. Kafri, and C. Neill managed the teams and development, and Y. Chen oversaw the project. H. Neven created an environment where this research was possible. All authors contributed to building the hardware and software, and writing the paper.
Acknowledgements
We thank Daniel Sank for valuable discussions and feedback on this manuscript.
Google Quantum AI and Collaborators
Alec Eickbusch, Nick Noll, Michael Newman, Matt McEwen, Leigh S. Martin, Alexis Morvan, Gabrielle Roberts, Chunyang Ding, Volodymyr Sivak, Nathan Lacroix, Laura De Lorenzo, Felix Borjans, Yan-Cheng Wei, Alexandre Bourassa, Rodrigo G. Cortiñas, Emma Rosenfeld, Vladislav D. Kurilovich, Andreas Bengtsson, Kevin J. Satzinger, Dmitry Abanin, Amira Abbas, Rajeev Acharya, Laleh Aghababaie Beni, Ross Alcaraz, Sayra Alcaraz, Trond I. Andersen, Benjamin Anker, Markus Ansmann, Frank Arute, Kunal Arya, Walt Askew, Nikita Astrakhantsev, Juan Atalaya, Lamia Ateshian, Christopher Ayala, Ryan Babbush, Akim Babenko, Brian Ballard, Joseph C. Bardin, Hector Bates, Majid Bigdeli Karimi, Alexander Bilmes, Simon Bilodeau, Chris Bohn, Pablo Bonilla, Jenna Bovaird, Dylan Bowers, Leon Brill, Peter Brooks, Michael Broughton, David A. Browne, Brett Buchea, Bob B. Buckley, Chris Bunch, Tim Burger, Brian Burkett, Jamal Busnaina, Nicholas Bushnell, Jonas Bylander, Anthony Cabrera, Juan Campero, Hung-Shen Chang, Gloria Chen, Larry Chen, Silas Chen, Federico Chianese, Ben Chiaro, Liang-Ying Chih, Agnetta Y. Cleland, Bryan Cochrane, Matt Cockrell, Josh Cogan, Tom Connolly, Harold Cook, William Courtney, Alexander L. Crook, Ben Curtin, Adam Dally, Martin Damyanov, Sayan Das, Dripto M. Debroy, Sean Demura, Himadri Dey, Leon Ding, Howard A. Dobbs, Ebru Dogan, Paul Donohoe, Ilya Drozdov, Andrew Dunsworth, Rasmus Edholm, Valerie Ehimhen, Aviv Moshe Elbag, Lior Ella, Mahmoud Elzouka, David Enriquez, Catherine Erickson, Lara Faroo, Parvin Fathi, Vinicius S. Ferreira, Marcos Flores, Leslie Flores Burgos, Sam Fontes, Ebrahim Forati, Jeremiah Ford, Brooks Foxen, Masaya Fukami, Alan Wing Lun Fung, Lenny Fuste, Suhas Ganjam, Gonzalo Garcia, Christopher Garrick, Robert Gasca, Helge Gehring, Robert Geiger, Élie Genois, William Giang, Dar Gilboa, James E. Goeders, Edward C. Gonzales, Sofia Gonzalez Garcia, Mike Gore, Raja Gosula, Stijn J. de Graaf, Alejandro Grajales Dau, Dietrich Graumann, Joel Grebel, Alex Greene, Jeremy Grill, Jonathan A. Gross, Jose Guerrero, Loïck Le Guevel, Marika Gugole, Simon Gustavsson, Steve Habegger, Tanner Hadick, Ali Hadjikhani, Michael C. Hamilton, Bianca Hanly, Monica Hansen, Matthew P. Harrigan, Patrick Harrington, Sean D. Harrington, Jeanne Hartshorn, Christian Heslin, Stephen Heslin, Paula Heu, Oscar Higgott, Reno Hiltermann, Devon Hollowood, Mike Hucka, Christopher Hudspeth, Ashley Huff, William J. Huggins, Lev B. Ioffe, Aditya Jayaraman, Evan Jeffrey, Shaun Jewons, Zhang Jiang, Xiaoxuan Jin, Tomas Jochym-O’Connor, Shunnee Joseph, Chaitali Joshi, Kelsey Josund, Kelsey Judd, Pavol Juhas, Andreas Kabel, Hui Kang, Kiseo Kang, Bharath Kannan, Amir H. Karamlou, Ryan Kaufman, Kostyantyn Kechedzhi, Tanuj Khattar, Mostafa Khezri, Seon Kim, Paul V. Klimov, Can M. Knaut, Tom Ko, Bryce Kobrin, Fedor Kostritsa, John Mark Kreikebaum, Ryuho Kudo, Ben Kueffner, Vitali Kutsko, Dylan Lane, Tianо Lange-Dei, Brandon W. Langley, Pavel Laptev, Kim-Ming Lau, Emma Leavell, Justin Ledford, Joy Lee, Kenny Lee, Nathalie de Leon, Brian J. Lester, Wendy Leung, Lily Li, Ming Li, Wing Yan Li, Alexander T. Lill, Mike Lindmark, William P. Livingston, Matthew T. Lloyd, Jing Lu, Erik Lucero, Daniel Lundahl, Aaron Lunt, Donald Macaskill, Sid Madhuk, Brian Maertz, Aniket Maitra, Ashley Maloney, Salvatore Mandrà, Bhavik Markana, Orion Martin, Eric Mascot, Paul Masih Das, Anselme Massinominol, Melvin Mathews, Cameron Maxfield, Jarrod R. McClean, Seneca Meeks, Anthony Megrant, Tim Menke, Zlatko K. Minev, Reza Molavi, Sebastian Molina, Shirin Montazeri, William Morong, Wojtek Muczkiewicz, Aki Nakamura, Konstantin Nesterov, Roselle Ngoly, Anthony Nguyen, Murray Nguyen, Chia-Hung Ni, Murphy Yuezhen Niu, Sergey Novikov, Ricky Oliver, William D. Oliver, Christopher Olson, Raymond Orosco, Amr Osman, Kristoffer Ottosson, Alice Pagano, Augustin Di Paolo, Seong Hyeon Park, Manan Patel, Sherman Peek, David Peterson, Alex Pizzo, Thomas Plumb-Reyes, Elias Portoles, Rebecca Potter, Orion Pritchard, Sam Probst, Michael Qian, Chris Quintana, Matthew Rakher, Ganesh Ramachandran, Arpit Ranadive, Armin Razavi, Matthew J. Reagor, Ants Remm, Rachel Resnick, Andrew Reynolds, David M. Rhodes, Daniel Riley, Roberto Rodriguez, Emma Ropes, Lucia B. De Rosel, Eliott Rosenberg, Dario Rosenstock, Elizabeth Rossi, Pedram Roushan, David A. Rower, Robert Salazar, Kannan Sankaragomathi, Murat Can Sarihan, Max Schaefer, Sebastian Schroeder, Henry F. Schurkus, Aria Shahingohar, Michael J. Shearn, Aaron Shorter, Vladimir
Corresponding authors
Correspondence and requests for materials should be addressed to Alec Eickbusch ([email protected]) and Nick Noll ([email protected]).
Competing Interests
The authors declare no competing interests.
Code and data availability
The datasets generated and analyzed for this study are available at https://doi.org/10.5281/zenodo.24142014. Shvarts{sup}1, Mauricio Sifontes{sup}1, Spencer Small{sup}1, W. Clarke Smith{sup}1, David A. Sobel{sup}1, Barrett Spells{sup}1, Sofia Springer{sup}1, Praveen Sriram{sup}1, George Sterling{sup}1, Scott Stonemeyer{sup}1, John Su{sup}1, Jordan Suchard{sup}1, Philippe Suchsland{sup}1, Youngkyu Sung{sup}1, Aaron Szasz{sup}1, Alexander Sztein{sup}1, Tomoyuki Tanaka{sup}1, Madeline Taylor{sup}1, Jothi Priyanka Thiruraman{sup}1, Douglas Thor{sup}1, Brandur Thorgrimsson{sup}1, Dogan Timucin{sup}1, Eifu Tomita{sup}1, Alfredo Torres{sup}1, M. Mert Torunbalci{sup}1, Edoardo Trabaldo{sup}1, Hao Tran{sup}1, Takahiro Tsunoda{sup}1, Justin Vargas{sup}1, Gopinath Vasalamarri{sup}1, Sergey Vdovichev{sup}1, Guifre Vidal{sup}1, Benjamin Villalonga{sup}1, Jeremy Volkman{sup}1, Meghan Voorhees{sup}1, Steven Waltman{sup}1, Jonathan Waltz{sup}1, Danni Wang{sup}1, Shannon X. Wang{sup}1, Brayden Ware{sup}1, James D. Watson{sup}1, Yonghua Wei{sup}1, Travis Weidel{sup}1, Tom Westerhout{sup}1, Theodore White{sup}1, Dominik Wild{sup}1, Kristi Wong{sup}1, Bryan W. K. Woo{sup}1, Christopher J. Wood{sup}1, Maddy Woodson{sup}1, Cheng Xing{sup}1, Z. Jamie Yao{sup}1, Ping Yeh{sup}1, Bicheng Ying{sup}1, Juhwan Yoo{sup}1, Noureldin Yosri{sup}1, Elliot Young{sup}1, Grayson Young{sup}1, Adam Zalcman{sup}1, Ran Zhang{sup}1, Yaxing Zhang{sup}1, Yiqing Zhou{sup}1, Ningfeng Zhu{sup}1, Nicholas Zobrist{sup}1, Zhenjie Zou{sup}1, Alexander N. Korotkov{sup}1, Sergio Boixo{sup}1, Hartmut Neven{sup}1, Kevin C. Miao{sup}1, Alex Opremcak{sup}1, Zijun Chen{sup}1, Cody Jones{sup}1, Julian Kelly{sup}1, Dvir Kafri{sup}1, Charles Neill{sup}1, Craig Gidney{sup}1, Yu Chen{sup}1`
{sup}1` Google Quantum AI, Santa Barbara, CA 93117, USA
{sup}2` Department of Electrical and Computer Engineering, University of Massachusetts, Amherst, MA, USA
{sup}3` Department of Physics, University of California, Santa Barbara, CA, USA
{sup}4` Department of Electrical and Computer Engineering, Auburn University, Auburn, AL, USA
{sup}5` Department of Physics, University of Connecticut, Storrs, CT, USA
{‡} These authors contributed equally to this work.
References
[1] A. Y. Kitaev, Fault-tolerant quantum computation by anyons, Annals of Physics 303, 2 (2003), arXiv:quant-ph/9707021 [quant-ph].
[2] A. G. Fowler, M. Mariantoni, J. M. Martinis, and A. N. Cleland, Surface codes: Towards practical large-scale quantum computation, Phys. Rev. A 86, 032324 (2012).
[3] B. M. Terhal, Quantum error correction for quantum memories, Rev. Mod. Phys. 87, 307 (2015).
[4] Google Quantum AI and Collaborators, R. Acharya, D. A. Abanin, L. Aghababaie-Beni, I. Aleiner, T. I. Andersen, M. Ansmann, F. Arute, K. Arya, A. Asfaw, N. Astrakhantsev, J. Atalaya, R. Babbush, D. Bacon, B. Ballard, J. C. Bardin, J. Bausch, A. Bengtsson, A. Bilmes, S. Blackwell, S. Boixo, G. Bortoli, A. Bourassa, J. Bovaird, L. Brill, M. Broughton, D. A. Browne, B. Buchea, B. B. Buckley, D. A. Buell, T. Burger, B. Burkett, N. Bushnell, A. Cabrera, J. Campero, H.-S. Chang, Y. Chen, Z. Chen, B. Chiaro, D. Chik, C. Chou, J. Claes, A. Y. Cleland, J. Cogan, R. Collins, P. Conner, W. Courtney, A. L. Crook, B. Curtin, S. Das, A. Davies, L. De Loronzo, D. M. Debroy, S. Demura, M. Devoret, A. Di Paolo, P. Donohoe, I. Drozdov, A. Dunsworth, C. Earle, T. Edlich, A. Eickbusch, A. M. Elbag, M. Elzouka, C. Erickson, L. Faoro, E. Farhi, V. S. Ferreira, L. Flores Burgos, E. Forati, A. G. Fowler, B. Foxen, S. Ganjam, G. Garcia, R. Gasca, E. Genois, W. Giang, C. Gidney, D. Gilboa, R. Gosula, A. Grajales Dau, D. Graumann, A. Greene, J. A. Gross, S. Habegger, J. Hall, M. C. Hamilton, M. Hansen, M. P. Harrigan, S. D. Harrington, F. J. H. Heras, S. Heslin, P. Heu, O. Higgot, G. Hill, J. Hilton, G. Holland, S. Hong, H.-Y. Huang, A. Huff, W. J. Huggins, L. B. Ioffe, S. V. Isakov, J. Iveland, E. Jeffrey, Z. Jiang, C. Jones, S. Jordan, C. Joshi, P. Juhas, D. Kafri, H. Kang, A. H. Karamlou, K. Kechedzhi, J. Kelly, T. Khaire, T. Khattar, M. Khezri, S. Kim, P. V. Klimov, A. R. Klots, B. Kobrin, P. Kohli, A. N. Korotkov, F. Kostritsa, R. Kothari, B. Kozlovskii, J. M. Kreikebaum, V. D. Kurilovich, N. Lacroix, D. Landhuis, T. Lange-Di, B. W. Langley, P. Laptev, K.-M. Lau, L. Le Guevel, J. Ledford, J. Lee, K. Lee, Y. D. Lensky, S. Leon, B. J. Lester, W. Y. Li, Y. Li, A. T. Lill, W. Liu, W. P. Livingston, A. Locharla, E. Lucero, D. Lundahl, A. Lunt, S. Madhuk, F. D. Malone, A. Maloney, S. Mandrà, J. Manyika, L. S. Martin, O. Martin, S. Martin, C. Maxfield, J. R. McClean, M. McEwen, S. Meeks, A. Megrant, X. Mi, K. C. Miao, A. Mieszala, R. Molavi, S. Molina, S. Montazeri, A. Morvan, R. Movassagh, W. Mruçzkiewicz, Q. Naaman, M. Neeley, C. Neill, A. Nersisyan, H. Neven, M. Newman, J. H. Ng, A. Nguyen, M. Nguyen, C.-H. Ni, M. Y. Niu, T. E. O’Brien, W. D. Oliver, A. Opremcak, K. Ottosson, A. Petukhov, A. Pizzuto, J. Platt, R. Potter, O. Pritchard, L. P. Pryadko, C. Quintana, G. Ramachandran, M. J. Reagor, J. Redding, D. M. Rhodes, G. Roberts, E. Rosenberg, E. Rosenfeld, P. Roushan, N. C. Rubin, N. Saei, D. Sank, K. Sankaragomathi, K. J. Satzinger, H. F. Schurkus, C. Schuster, A. W. Senior, M. J. Shearn, A. Shorter, N. Shutty, V. Shvarts, S. Singh, V. Sivak, J. Skrzuny, S. Small, V. Smelyanskiy, W. C. Smith, R. D. Somma, S. Springer, G. Sterling, D. Strain, J. Suchard, A. Szasz, A. Sztein, D. Thor, A. Torres, M. M. Torunbalci, A. Vaishnav, J. Vargas, S. Vdovichev, G. Vidal, B. Villalonga, C. Vollgraff Heidweiller, S. Waltman, S. X. Wang, B. Ware, K. Weber, T. Weidel, T. White, K. Wong, B. W. K. Woo, C. Xing, Z. J. Yao, P. Yeh, B. Ying, J. Yoo, N. Yosri, G. Young, A. Zalcman, Y. Zhang, N. Zhu, and N. Zobrist, Quantum error correction below the surface code threshold, Nature 638, 920 (2025), arXiv:2408.13687 [quant-ph].
[5] C. Gidney, Stability experiments: The overlooked dual of memory experiments, Quantum 6, 786 (2022).
[6] P. W. Shor, Scheme for reducing decoherence in quantum computer memory, Phys. Rev. A 52, R2493 (1995).
[7] S. Krinner, N. Lacroix, A. Remm, A. Di Paolo, É. Genois, C. Leroux, C. Hellings, S. Lazar, F. Swiadek, J. Herrmann, G. J. Norris, C. K. Andersen, M. Müller, A. Blais, C. Eichler, and A. Wallraff, Realizing repeated quantum error correction in a distance-three surface code, Nature 605, 669 (2022), arXiv:2112.03708 [quant-ph].
[8] Google Quantum AI, R. Acharya, I. Aleiner, R. Allen, T. I. Andersen, M. Ansmann, F. Arute, K. Arya, A. Asfaw, J. Atalaya, R. Babbush, D. Bacon, J. C. Bardin, J. Basso, A. Bengtsson, S. Boixo, G. Bortoli, A. Bourassa, J. Bovaird, L. Brill, M. Broughton, B. B. Buckley, D. A. Buell, T. Burger, B. Burkett, N. Bushnell, Y. Chen, Z. Chen, B. Chiaro,
Stability experiment layout and definitions
A stability experiment encodes zero logical qubits and tracks a global stabilizer product across time to isolate timelike measurement and ancilla errors. Surrounding a patch entirely with boundaries of one type ( or ) cancels all data-qubit operators in the product of stabilizers of that type (), defining a global parity observable that is constant across error-free cycles. This can be seen by noticing that all data qubits enter into the product an even number of times.
Each stability patch is specified by its width , a fixed bulk checkerboard tiling, and the boundary basis ( or ; Figure S1). Defining the bulk tiling first locks the four-layer CZ gate order across -stability, -stability, and paired memory runs, after which choosing the basis keeps only the matching weight-2 perimeter checks. Wherever a corner bulk tile matches the boundary basis, the adjacent perimeter checks are absent, so that corner data qubit belongs to only one check in and is removed to leave a weight-3 corner check. Even widths () therefore keep all four corners in one basis and cut all four in the other, whereas odd widths () cut opposite pairs of corners.
Figure S1. Stability patch construction across width and basis. Patches for (a–c) and (d–f) on a fixed bulk checkerboard. a, d, Bulk tiling with interior arrows showing the two-qubit CZ gate order and dashed arcs showing candidate weight-2 boundaries. b, e, Basis X and c, f, Basis Z patches; green circles mark measure qubits included in , and dashed circles mark removed corner data qubits.
Data qubits are initialized in and measured in the basis in the first and last cycles. Since single-qubit states anticommute with individual and checks, all individual initial and final stabilizer outcomes are randomized while their global product remains deterministic. Before the measure-qubit readout Hadamard in each cycle, virtual gates (s in Figure S2) are randomly applied, realizing random per-shot sign flips to the ancillas in . Keeping these classical bits fixed across cycles leaves two-cycle detectors unchanged while randomizing from shot to shot.
Figure S2. Single-cycle M-to-M detector slice of the ZXXZ stability circuit. Steady-state cycle for the , Basis patch of Figure S1 (blue: sensitivity; red: sensitivity). Ancilla measurements (a, l) close and reopen the ZXXZ detector slice; reset and Hadamard (b, c) precede four CZ layers (d, f, g, i) interleaved with data Hadamards (e, h); and classical-bit-controlled virtual gates (, j) precede the readout Hadamards (k).
We run paired memory and stability experiments across 15 subpatches spanning on a 105-qubit superconducting processor (Figure S3). At , 4, and 5 (Figure S3–Figure S3), subpatches sit away from the left and right chip edges so that all perimeter ancillas exist, allowing Memory ( and ; white), Stability (red), and Stability (green) to run on every location. At and (Figure S3,Figure S3), peripheral ancillas on the top and bottom boundaries support Stability alongside Memory and , whereas the left and right edges carry no extra -boundary ancillas, so Stability is omitted. On hardware, a few additional neighboring qubits outside each patch are also used for data-qubit leakage removal (DQLR) on boundary data qubits.
Figure S3. Memory and stability subpatch tilings on a 105-qubit superconducting processor. Outlines show Memory ( and ; white), Stability (red), and Stability (green) across a, seven , b, four , c, two , d, one , and e, one full-chip subpatch (gray crosses mark unused peripheral qubits; darker and lighter blue tiles show local and plaquette sublattices).
Heuristic model including temporally correlated errors
In a stability experiment of cycles, a timelike logical error occurs when errors accumulate on at least cycles along a timelike path, causing a minimum-weight decoder to match to the and time boundaries instead of pairing errors locally. With uncorrelated Pauli errors alone, the logical error probability decays exponentially with as , where is the per-cycle Pauli suppression factor and captures boundary effects and spatial path counting across the patch.
To model the knee at larger , suppose a rare correlated error occurs anywhere in the patch with per-shot probability (and does not occur with probability ), producing a contiguous timelike error chain of length cycles. For physical mechanisms with memory that have a constant probability per cycle of relaxing or being reset, such as leakage to higher transmon states or coupling to a two-level system (TLS) with a finite lifetime , the chain length follows an exponential distribution,
where is the characteristic correlation time in cycles and (since counts only the correlated error chains, while is the probability of having none).
If a chain triggers detection events only at its start and end cycles, any single chain with spans half the experiment and causes a logical error on its own. Neglecting combinations of shorter chains with Pauli errors () and taking even for simplicity, summing over gives . Adding this to the Pauli term gives the heuristic sum-of-exponentials model used in the main text,
with
so for . Whenever correlated chains decay more slowly than Pauli errors (), bends at a knee from steep Pauli suppression to a shallower tail set by .
In practice, a leaked qubit or active TLS can also trigger detection events during the middle of the chain with some probability per cycle, allowing the decoder to pair segments locally in the bulk rather than to the time boundaries. In addition, since this simple picture also leaves out starting-time multiplicity, hybrid Pauli-correlated chains, and three-dimensional decoding details, we treat Eq. (2) as a phenomenological fit and use circuit-level simulations to study specific error mechanisms.
Chain-length distributions, staircases, and multiple knees
An exponential chain-length distribution is most consistent with leakage and other physical temporally correlated error mechanisms with memory, such as coupling to a TLS with a finite lifetime , and produces a single smooth decay after the knee. Different correlated error processes, however, can have different chain-length distributions , producing staircases, regions where the logical error worsens with cycles, or multiple knees (Figure S4).
Figure S4. Simulated stability experiments under different correlated error models. Top row: correlated chain-length probability per stabilizer and per cycle versus chain length , with . Bottom row: simulated stability logical error probability versus cycles for a -basis surface code patch under SI1000 noise (), decoded with correlated minimum-weight perfect matching. Correlated errors are added after circuit sampling by flipping a stabilizer’s measurement outcomes for consecutive cycles. a) Fixed-length chains, , for (, orange squares) and (, blue circles). Within each step, rises with because a longer experiment provides more opportunities in time for a limiting error chain to occur, before dropping sharply at multiples of once an additional chain is needed to span half the experiment. b) Exponentially distributed chains, , consistent with leakage or coupling to a TLS with finite lifetime , shown for (red squares) and (green triangles). Both decay smoothly after the knee and cross near , where the longer-lived error takes over. c) Combined models with multiple knees: combining two exponentials with and (cyan triangles) produces two knees separating three monotonic decay slopes, while combining fixed-length chains with an exponential tail (purple circles) produces a rising plateau up to followed by a drop into the tail.
To build intuition for these features, we simulate a stability experiment and inject correlated errors by flipping a stabilizer’s measurement outcomes for consecutive cycles with probability per cycle. When chains have a fixed length (Figure S4), a single chain is long enough to cause a logical error whenever . Within this window, running the experiment for more cycles gives more opportunities in time for such a limiting error to occur, a temporal path-counting effect analogous to the spatial path counting in the main text, causing to rise with . Once , a single chain is too short to span half the experiment on its own and two chains are required, causing a sharp drop onto the next step and forming a staircase with period . For an exponential distribution (Figure S4), longer chains are exponentially rarer, overcoming path counting to give a smooth post-knee decay where longer-lived errors eventually overtake shorter-lived ones at large . Combining mechanisms (Figure S4) produces multiple knees, either bending smoothly between distinct exponential slopes or rising along a step before dropping into an exponential tail.
Numerical model
To capture both stochastic Pauli errors and multi-cycle temporal correlations arising from higher excited transmon states, we simulate our experiments using two complementary frameworks: (i) a circuit-level stochastic Pauli model (implemented in Stim [47]) for rapid detector statistics and Pauli-regime error budgeting, and (ii) a multi-level stabilizer-frame simulator tracking states , , , and for full leakage dynamics.
During data-qubit idling and dynamical decoupling (DD) windows (which coincide with measure-qubit readout and reset), we model incoherent population transfer among the lowest four transmon levels – as a continuous-time Markov process parameterized by empirical transition times . Fundamental relaxation () and dephasing () rates are assigned per qubit from individual idle and DD benchmarks, while higher-level relaxation times (, ), single-photon heating times (, , ), and direct two-photon heating times (, ) are applied uniformly across the grid based on device characterization. At the end of each syndrome extraction cycle, Data Qubit Leakage Removal (DQLR) [9] is applied to all data qubits.
We model DQLR as a state-selective stochastic reset channel on leaked data qubits that returns with probability , while clearing with probability , set to the median characterized values from the device. Measure qubits undergo a multi-level reset each cycle that clears both and back to the computational subspace, up to the benchmarked reset error . We assume all leaked states are read out as “1” in the simulation.
The residence time of a data qubit in is governed by the interplay between intrinsic relaxation () and population transfer during two-qubit CZ gates (“leakage transport”). In order to model multi-level leakage transport during CZ gates accurately without full waveform-level master-equation integration across the grid, we construct an effective block-diagonal unitary channel for each interacting qubit pair from a coupled transmon–coupler–transmon Duffing oscillator model parameterized by the calibrated idle frequencies and measured anharmonicities of both qubits, along with the calibrated interaction frequency and hold time of the gate. Leakage transport is mediated when a data qubit in participates in a CZ gate with a measure qubit in ( manifold spanned by ) or ( manifold spanned by ): Bose-enhanced off-resonant exchange transfers excitation from the data qubit onto the measure qubit (e.g., or ). As each bulk data qubit executes four CZ gates per cycle with neighboring measure qubits (which are reset every cycle) and undergoes DQLR (which clears any resulting population on the data qubit with 98% probability), this CZ-mediated transport provides the primary removal pathway for data-qubit excitations, setting the post-knee stability decay rate observed in Figure 3a.
Component error rates
To calibrate the circuit-level Pauli noise model used in our surface code simulations, we benchmark single-qubit (1Q) and two-qubit (2Q) operations in parallel circuit layers matching the syndrome-extraction schedule. Figure S5 shows the empirical cumulative distribution functions (ECDFs) and boxplots of the directly measured component error rates across the Willow processor, and Table S1 summarizes the mean, median, and standard deviation of each distribution. All error rates are reported in units of Pauli error probability , defined as the total probability of any non-identity Pauli error (related to the average gate infidelity in Hilbert-space dimension via ).
Figure S5. Distributions of directly measured component Pauli error probabilities entering the surface code simulation. (a) Empirical cumulative distribution functions (ECDFs) of the per-target Pauli error probabilities across the processor; circular markers along the dotted line (ECDF = 0.5) indicate the median of each distribution. The noise floor of the reset benchmark caps the distribution at Pauli error . (b) Corresponding boxplots for each component benchmark on a logarithmic vertical scale, showing the interquartile range (25th–75th percentiles, shaded boxes), median (horizontal solid line), 5th–95th percentile range (whiskers), mean (open diamonds), and individual target measurements (points).
| Component | Mean | Median | Std. Dev. | |
| Reset error () | 52 | 0.125% | 0.063% | 0.257% |
| 1Q gate error () | 104 | 0.086% | 0.067% | 0.058% |
| Data-qubit leakage removal (DQLR) error () | 49 | 0.211% | 0.172% | 0.122% |
| 2Q XEB cycle error () | 180 | 0.454% | 0.380% | 0.230% |
| Readout error () | 101 | 0.938% | 0.757% | 0.524% |
| Dynamical decoupling (DD) idle error () | 55 | 1.115% | 0.816% | 0.874% |
Table S1. Mean, median, and standard deviation of the directly measured component Pauli error probabilities (%) across benchmarked targets entering the surface code simulation noise model.
Specifically, simultaneous 1Q Clifford randomized benchmarking (RB) measures the 1Q gate infidelity (), while parallel 2Q cross-entropy benchmarking (XEB) measures the cycle infidelity of two parallel 1Q gates followed by a controlled-Z (CZ) gate (; in circuit simulations where 1Q gate errors are applied separately, the isolated CZ gate error is obtained by subtracting the two constituent 1Q gate errors, ). Interleaved 1Q RB directly measures the Pauli error of the data-qubit leakage removal (DQLR) operation, and state-preparation-and-measurement benchmarks characterize the residual excited-state probability after reset () and the state-averaged readout assignment error (), both modeled as bit-flip () errors. Finally, while ancilla qubits undergo measurement and reset, data qubits idle under dynamical decoupling (DD), where longitudinal and transverse purity decay benchmarks measure the per-cycle relaxation and dephasing probabilities and (giving a total Pauli error under Pauli twirling).
Error budgeting
To construct the Pauli-regime error budgets shown in Figure 2e,Figure 2f, we simulate memory, stability, and combined spacetime logical error probabilities across odd code distances using our benchmarked circuit-level Pauli noise model (§ 1.16) and ablate each error channel family one at a time. The optimal memory error , stability error , and minimum spacetime error are evaluated on a spacetime block () at the fixed experimental optimal cycle counts for (held constant across the baseline and all single-channel ablations, using the multi-cycle curve fits to suppress single-cycle Monte Carlo statistical fluctuations), where stability accounts for 16% of the combined logical error compared to 84% from memory ( ratio in Figure 2f; Table S3). Fitting across yields the baseline inverse block suppression factors and single-channel ablated values . Table S2 reports the resulting channel sensitivities , normalized budget shares , and stability-to-memory relative contribution ratios .
Table S2. Space (), Time (), and Spacetime () block suppression error budgets fitted over odd distances () at fixed optimal cycle count . Entries give sensitivity and relative budget share (%). Propagated shot-noise uncertainties on are – (< 1% of budget) and are shown explicitly for the Stability/Memory () relative contribution ratio.
| Channel | Memory () | Stability () | Spacetime () | Ratio () |
| 2Q Gates | 0.336 (55.3%) | 0.291 (43.4%) | 0.327 (54.1%) | 0.78 ± 0.02× |
| Idle | 0.193 (31.7%) | 0.165 (24.5%) | 0.189 (31.3%) | 0.77 ± 0.02× |
| Readout | 0.030 (4.9%) | 0.141 (21.0%) | 0.037 (6.2%) | 4.28 ± 0.27× |
| 1Q Gates | 0.024 (4.0%) | 0.028 (4.2%) | 0.025 (4.1%) | 1.05 ± 0.12× |
| Reset | 0.015 (2.5%) | 0.041 (6.0%) | 0.017 (2.7%) | 2.43 ± 0.34× |
| DQLR | 0.010 (1.7%) | 0.006 (0.9%) | 0.009 (1.5%) | 0.53 ± 0.27× |
| Sum | 0.608 (100%) | 0.672 (100%) | 0.605 (100%) | — |
Table S2.
Table S3. Baseline block logical error probability () and single-channel error reductions () evaluated at the fixed optimal cycle count across odd distances .
| Metric | Full | 2Q | Idle | RO | 1Q | Reset | DQLR | |
| Memory | 3 () | 58.8 | 31.9 | 31.3 | 15.8 | 3.0 | 1.2 | 1.3 |
| 5 () | 43.3 | 33.3 | 29.9 | 12.1 | 3.6 | 2.2 | 1.6 | |
| 7 () | 22.8 | 20.6 | 17.7 | 7.7 | 2.8 | 1.5 | 1.2 | |
| Stability | 3 () | 14.5 | 8.3 | 4.3 | 8.4 | 1.0 | 3.7 | 0.1 |
| 5 () | 5.4 | 4.5 | 2.9 | 3.7 | 0.7 | 1.6 | 0.1 | |
| 7 () | 4.7 | 4.2 | 3.0 | 3.6 | 0.7 | 1.7 | 0.1 | |
| Spacetime | 3 () | 36.6 | 20.1 | 17.8 | 12.1 | 2.0 | 2.4 | 0.7 |
| 5 () | 24.3 | 18.9 | 16.4 | 7.9 | 2.2 | 1.9 | 0.8 | |
| 7 () | 13.7 | 12.4 | 10.4 | 5.6 | 1.8 | 1.6 | 0.7 |
Table S3.
Complementary gap
Decoding details
We decode the surface code stability experiments using a superconducting-inspired (SI1000) circuit-level Detector Error Model (DEM) prior coupled with a correlated minimum-weight perfect matching (MWPM) decoder [39]. The SI1000 noise model [40] is parameterized by a single characteristic physical error probability , assigning two-qubit CZ gate depolarizing error , single-qubit gate depolarizing error , data-qubit idle error during gate layers, measurement error , reset bit-flip probability , and spectator data-qubit idle error during measurement and reset windows.
Definition and computation
For a single experimental shot, let be the total log-likelihood weight of the minimum-weight error configuration satisfying the syndrome constraints. Let denote the weight of the minimum-weight error configuration consistent with the same syndrome but constrained to flip the logical observable to the opposite class. The complementary gap is defined as the log-likelihood weight difference
and the signed complementary gap relative to the true stability observable is
In our pipeline, is computed at each candidate prefix window ( for the static surface code, with at and at ; and even cycles with for the two-cycle-period walking surface code) using a fixed logical boundary detector and two matching runs—once with the detector on and once with the detector off [36].
Adaptive stopping rules
In an online adaptive experiment, the controller observes only the unsigned gap and the syndrome history up to cycle , never the sign of . On most shots, errors are independent Pauli faults and grows steadily in the correct logical class, allowing the experiment to terminate as soon as crosses a confidence threshold . During a multi-cycle leakage episode, however, the signed gap can sometimes grow steadily negative—meaning the decoder becomes increasingly confident in the wrong logical outcome. This happens because the decoder assumes all faults are independent: when a leaked qubit deposits a stream of detection events that locally pair up like natural, independent single-cycle errors, it can appear much cheaper to connect an initial defect to the early time boundary and treat the subsequent detections as independent local pairs than to track a single persistent fault across the entire window. Only after the leaked qubit relaxes back to the computational subspace does the true explanation become favored again, causing to recover and grow positive. Depending on whether the circuit architecture suppresses stationary multi-cycle leakage streaks, we employ two causal adaptive stopping strategies:
Pure gap stopping (used for the walking surface code): A shot stops at the first candidate cycle where the complementary gap exceeds a fixed threshold,
Because the walking surface code swaps data- and measure-qubit roles each cycle and applies echo gates, it is more difficult for a leaked qubit to sit on a fixed site and repeatedly deposit deceptive local detection pairs on the same checks; consequently, pure gap stopping suffices without an auxiliary syndrome filter, which can unnecessarily inflate the mean stopping time.
Syndrome-gated gap stopping (used for the standard surface code): To prevent premature termination while a stationary leaked qubit is driving negative, we pair the gap threshold with a causal local syndrome-burst indicator . At each detector round , if any data qubit’s trailing -cycle neighborhood detection count exceeds a binomial -score threshold above its baseline firing rate. A shot stops at the first cycle where the neighborhood is currently quiet and the gap exceeds a threshold raised by the cumulative number of previously flagged cycles :
Holding the shot while prevents stopping mid-burst, while increasing the threshold by nats per flagged cycle ensures that once the burst ends, the experiment waits enough additional cycles for to recover from negative values back above . In our analyses of the static and patches, each data qubit’s neighborhood comprises its adjacent measure qubits tracked over a trailing window of detector rounds, with baseline per-detector firing probability estimated from the empirical ensemble mean and burst threshold set to . The threshold-inflation slope is set to nats per flagged cycle for the static patch (matching the mean Pauli gap growth rate) and nats per flagged cycle for the patch; any shot not meeting the stopping criterion before the end of the recorded window terminates at . To generate the adaptive tradeoff curves in Figure 5c,Figure 5d, we sweep nats; for the spacetime error minimum in Figure 5d ( nats), the hyperparameters are selected in-sample on the hardware dataset, with out-of-sample generalization verified via a 50/50 train/test split across shots.
Enhanced adaptive gains under time-correlated errors
While adaptive stopping already reduces the mean cycle count on Pauli-dominated shots by terminating high-confidence shots early, its advantage grows much larger in the presence of rarer, longer-lived time-correlated error mechanisms. The core intuition is simple: a fixed-cycle protocol must apply the same duration to every shot, forcing a tradeoff between stopping early on a correlated error floor or waiting long enough to catch rare bursts while accumulating extra spacelike memory error on every clean shot. Adaptive stopping breaks this tradeoff by finishing clean shots early and paying the long wait time only on the rare shots where a burst actually occurs.
To illustrate this effect in a toy model, we extend the two-component stability decay to include a hierarchy of correlated error tiers with steady-state patch probability and correlation lifetime :
Here tier 1 ( cycles, ) matches the measured -leakage floor on the patch, while hypothetical tiers occur less often at each step ( and at ) and persist twice as long ( and cycles), producing a staircase of flatter error shelves (Figure S6)—as could arise from excitations into progressively higher transmon states that cascade back down to the computational subspace.
Figure S6. Adaptive stopping bypasses tiered time-correlated error shelves in a toy model. a, Stability logical error versus cycles at . The Pauli decay (gray dashed) and first correlated shelf (, 10.5 cycles; amber dashed) are matched to experiment, while two rarer hypothetical correlated mechanisms (, 21 cycles, orange dashed; , 42 cycles, crimson dashed) have 10× smaller probability at each step and double in lifetime, producing a staircase of flatter error shelves in the total curve (solid black). b, Spacetime error versus code distance for a hypothetical , with optimal adaptive mean cycle count cycles indicated on the horizontal axis. The Pauli-only adaptive lower bound (dashed light-blue) scales the optimal fixed-Pauli spacetime error at (dashed gray) by the experimental early-stopping ratio at the same timelike Pauli error. Under a per-cycle patch burst rate , optimizing the adaptive stopping window to wait out rare bursts lifts the adaptive strategy with non-Pauli error (blue squares) by at most 1.5× above the Pauli-only adaptive baseline. By contrast, a fixed-cycle strategy evaluated at the same mean cycle count (black diamonds) transitions from Pauli-limited at to being pinned on the first residual shelf across – (with shelves 2 and 3 lying immediately below).
To project spacetime scaling across , we pair this stability decay with a hypothetical per-cycle memory error rate . Under adaptive stopping, burst-free shots terminate early at cycles, while shots flagged on tier are extended by cycles until the burst clears and its residual timelike error drops to the spacelike error scale:
Because a tier- burst has steady-state patch probability and lifetime , bursts initiate across the patch at rate per cycle. The probability of flagging a shot during the unflagged window therefore scales linearly with its spacetime volume . Even at , over 95% of shots finish unflagged, keeping the optimal mean runtime modest ( cycles for ) and holding the adaptive spacetime error within of the Pauli-only adaptive baseline (Figure S6).
By contrast, a fixed-cycle strategy given the same average cycle budget never waits long enough ( cycles) to get past the first correlated shelf, stalling near across – (Figure S6). At matched mean runtime, adaptive stopping yields a spacetime error advantage that grows substantially from to . Even if the fixed cycle count is instead globally increased to wait past the shelves (up to cycles at ), adaptive stopping still achieves – lower spacetime error while using – fewer cycles on average.
Walking Stability Experiment
In a standard surface code stability experiment, physical qubits are permanently assigned as either data qubits or measure qubits. In the walking stability experiment, the code patch instead translates back and forth across a grid of 25 physical qubits by swapping the roles of data and measure qubits in every cycle (Figure S7). During each cycle’s entangling layers, the stabilizers on the active data qubits are transferred onto the newly reset measure qubits before the old data qubits are measured and reset. Over two consecutive cycles, the patch walks by one lattice site from data qubits to measure qubits and then walks back to its original position. A stability experiment requires that the product of all stabilizers of the boundary type equals , so that their combined syndrome outcomes form a deterministic global parity check in the absence of errors. In a static stability patch, this constraint is satisfied using weight-4 bulk stabilizers surrounded by weight-2 boundary stabilizers. In the walking circuit, as the patch shifts across the fixed qubit grid, the boundaries at the end of each cycle include weight-3 stabilizers paired with weight-1 single-qubit stabilizers/detectors on the outer boundary qubits (Figure S7). Combining each weight-3 boundary stabilizer with its adjacent weight-1 boundary detector cancels the outer boundary qubit to form an effective weight-2 boundary stabilizer. Because every active data qubit is included an even number of times (twice) across the boundary-type stabilizers and detectors, their overall product remains throughout the walk.

Figure S7. Detector slices over two error-correction cycles of the walking stability experiment. Stim detector-slice diagram (detslice-with-ops, read left-to-right, top-to-bottom) beginning and ending at a measurement layer () with two full syndrome-extraction cycles in between. Colored regions indicate the instantaneous Pauli support of each active detector ( in red, in green, in blue). In the first cycle, the patch walks from the initial data sublattice onto the measure sublattice (middle measurement layer ); in the second cycle, it walks back to the original sublattice (final measurement layer ). At the measurement layers at the end of each cycle, weight-3 triangular boundary stabilizers pair with weight-1 single-qubit boundary detectors (blue circles on the boundary) to form effective weight-2 boundary stabilizers, preserving the product of all boundary-type stabilizers.
To compare the baseline fault-tolerant performance of the walking circuit against a static patch, we simulate both the walking stability circuit and a standard stability circuit under the SI1000 circuit-level noise model at a physical error rate (Figure S8). Each circuit is simulated for 30 repeatable error-correction cycles ( Monte Carlo shots) and analyzed using non-overlapping stability subsampling and minimum-weight perfect matching decoding. Both circuits exhibit exponential suppression of the logical error probability with cycle count , , with per-cycle suppression factors () for the standard stability circuit and () for the walking stability circuit. Because translating the patch across the grid and forming the weight-3 and weight-1 boundary detectors requires 25 physical qubits—compared to 15 physical qubits for the static stability circuit—the walking circuit incurs additional qubit and gate overhead, which accounts for the corresponding reduction in the error-suppression factor .
Figure S8. Pauli model simulation of 3 × 3 walking stability versus standard 3 × 3 stability. Logical error probability as a function of the number of error-correction cycles (up to 20 cycles) under the SI1000 circuit-level noise model with . Points are extracted via non-overlapping subsampling from 30-cycle Stim simulations ( base shots) and decoded with Sparse Blossom. Error bars indicate binomial confidence intervals (sinter, likelihood ratio factor ; vertical bars without markers at cycles indicate zero-event upper bounds). Solid lines show exponential suppression fits , yielding () for the standard 3 × 3 stability circuit (15 physical qubits) and () for the walking 3 × 3 stability circuit (25 physical qubits).
Example 20-Cycle 3 × 3 Walking Stability Stim Circuit
This crumble link is the complete Stim circuit for a 20-cycle 3 × 3 walking stability experiment (2 initialization cycles, 7 repetitions of the 2-cycle bulk loop, and 4 final cycles).
Stability Experiment Subsampling
The bulk of an -cycle stability experiment consists of identical, repeated syndrome-extraction cycles. As such, a single long run naturally contains the measurement history of shorter stability experiments within its interior. Rather than executing separate experiments at every cycle duration, we use subsampling to extract multiple shorter stability experiments from one long run (Figure S9).
Figure S9. Subsampling a stability experiment. (a) Spacetime volume of an -cycle stability experiment (beige), bounded by initial and final spacelike stabilizer planes (green) and a timelike stability observable (vertical dashed line). The bulk measurement record is sliced into windows (light blue) of duration span cycles separated by stride cycles. (b) Each window is decoded independently as a span-cycle stability experiment bounded by its first and last syndrome layers (orange planes).
In post-processing, we slice the bulk measurement record into smaller windows separated by a start-to-start offset stride (Figure S9). The first and last syndrome layers of each window define its temporal boundaries (Figure S9); detectors referencing earlier measurements are dropped, and each window is decoded independently. Sweeping window size across the same dataset yields the full logical error probability versus cycle curve from a single experiment.
Extracting windows from repetitions gives subsampled shots. For non-overlapping windows (stride span), all samples are independent. When windows partially overlap (stride span), shared measurements introduce correlations; we account for this by scaling down the raw shot count by the fraction of unique cycles covered (total distinct cycles spanned across all windows divided by ) and computing binomial error bars from the resulting effective shot count . For the walking experiment data in Figure 4a, all windows are independent.
We validate subsampling against direct stability experiments swept across cycle counts on 15 surface code patches (–). Subsampled curves are extracted solely from the -cycle run of each patch and decoded with the same settings as the direct runs. Error bars and shaded bands in Figure S10 combine patch-to-patch standard errors with the overlap-scaled binomial uncertainty from .
Figure S10. Experimental validation of stability subsampling across code distances –. (a) Patch-averaged logical error probability versus cycles for direct runs (open circles) and subsampled windows (solid curves, bands from ) from the -cycle run. Subsampling slightly over-predicts the error rate while accurately capturing the knee (gray band) and post-knee tail. (b) Two-exponential fits for comparing subsampled (solid) and direct (dashed) curves, with correlated-error components (dotted) and knee crossovers (stars). (c) Fitted pre-knee (, red) and post-knee (, blue) suppression factors versus code distance for direct (open circles) and subsampled (filled squares) data.
Across all five code distances, subsampling closely tracks the direct experiments with a small upward offset (Figure S10). Two effects make the subsampled error rate slightly higher: (1) temporal edge effects, because subsampled windows terminate on truncated bulk syndrome layers (orange planes in Figure S9) rather than physical state preparation and destructive data-qubit readout (green planes in Figure S9); and (2) steady-state leakage, which ramps up over the first handful of cycles—direct runs always start from a freshly reset state on all data and measure qubits, whereas subsampled windows are drawn from the interior of a 240-cycle run after leakage has equilibrated. Subsampling thus provides a slightly conservative upper bound on performance.
Crucially, subsampling accurately captures both the knee and the correlated-error regime beyond it. Using the two-exponential fit from the main text (Figure S10), subsampling reproduces the knee location (–) to within 1–2 cycles and recovers both the pre-knee Pauli suppression factor and the post-knee correlated-error suppression factor across – (agreeing within for ; Figure S10), while reducing statistical scatter in the low-error tail.
Reinforcement learning calibration
We use reinforcement learning (RL) to fine-tune calibration parameters for QEC, as described in Ref. [42]. In each RL epoch, we acquire QEC syndrome data from four circuits: distance- -basis memory, distance- -basis stability, and distance- -basis stability centered in two different locations. These circuits are selected to provide sufficient qubit coverage and ensure that control parameters are not overfit to any one circuit at the expense of losing performance in other circuits. We tune approximately 5,000 system control parameters simultaneously, including those described in Ref. [42] and additional new parameters. The error detection probability averaged across all detectors in these circuits decreases steadily during the course of training, as shown in Figure S11. We save the QEC syndrome data obtained with the fine-tuned RL policy for further post-processing and the analysis shown in the main text.

Figure S11. Evolution of average error detection probability during RL training on four QEC circuits.
Extended Data
In the main text, logical error probabilities are averaged across processor subpatches and across the X and Z Pauli bases. Here we present the stability and surface code memory data resolved by individual subpatch and basis (X, Z, and their combined mean) across four hardware runs.
Table S4 summarizes the consolidated shot counts across the four runs. Surface code memory is measured in both the X and Z bases across 24 cycle counts at all 15 subpatches ( shots per run and cycle count). Stability is measured in both bases for , , and , and in the Z basis only at and due to the processor boundary. Each stability run executes shots of a 200-cycle circuit; following Sec. S8, we subsample 5 evenly spaced, centered windows per shot within steady-state cycles at each of 19 window spans .
| Grids | Stability () | Memory (X, Z; ) | ||||
| Bases | Raw shots | Win./span | Shots/cyc. | Total (24 cyc.) | ||
| 3 | 7 | , | 28.0 | 140.0 | 2.80 | 67.2 |
| 4 | 4 | , | 16.0 | 80.0 | 1.60 | 38.4 |
| 5 | 2 | , | 8.0 | 40.0 | 0.80 | 19.2 |
| 6* | 1 | 2.0 | 10.0 | 0.40 | 9.6 | |
| 7* | 1 | 2.0 | 10.0 | 0.40 | 9.6 | |
| Total | 15 | 28 | 56.0 | 280.0 | 6.00 | 144.0 |
Table S4. Consolidated hardware shot counts and subsampled stability windows across four runs. Totals are in units of . Per subpatch and measured basis across the four runs, stability uses 2.0 raw 200-cycle shots (10.0 windows/span across 19 spans , 5 windows/shot) and memory uses 0.20 shots/cycle across 24 cycle counts in both and .
Extended Data Fig. S12 shows the stability logical error probability for each subpatch and basis [panels (a–e)], the subpatch-averaged two-component fits [panel (f)], the error across subpatches [panel (g) and Table S5], and the fast () and slow () suppression factors [panel (h)]. Every subpatch exhibits a rapid initial syndrome-suppression decay followed by a slower leakage-limited tail, with similar X and Z performance at odd patch widths and higher X-basis error at due to its larger boundary footprint. We note that the X and Z basis stability experiments have primary sensitivity to different sub-lattices of measure qubits. As a result, measure qubit outliers (in leakage or Pauli error rates) can cause one basis to degrade while not significantly impacting the other.
| Grids | Stability () | Memory /cyc. () | |||||
| Mean | Mean | ||||||
| 3 | 7 | 1.47 0.01 | 1.39 0.01 | 1.43 0.01 | 6.58 0.01 | 7.99 0.01 | 7.24 0.01 |
| 4 | 4 | 1.11 0.01 | 2.17 0.01 | 1.64 0.01 | 5.09 0.01 | 5.92 0.01 | 5.49 0.01 |
| 5 | 2 | 1.23 0.01 | 1.24 0.01 | 1.23 0.01 | 3.06 0.01 | 3.51 0.01 | 3.28 0.01 |
| 6* | 1 | 0.90 0.01 | — | 0.90 0.01 | 2.15 0.01 | 2.49 0.01 | 2.32 0.01 |
| 7* | 1 | 0.89 0.01 | — | 0.89 0.01 | 1.72 0.01 | 1.99 0.01 | 1.85 0.01 |
Table S5. Basis-resolved stability and memory summary metrics. Uncertainties denote 1 standard errors. At 6 6* and 7 7*, stability is measured in the basis only.
Extended Data Fig. S13 shows the corresponding surface code memory logical fidelity across individual subpatches [panels (a–e)], the subpatch-averaged exponential fidelity fits [panel (f)], and the per-cycle logical error rates [panel (g) and Table S5]. Across both experiments, changes in stability and memory performance across subpatches and bases are expected from qubit-to-qubit variation in both leakage rates and Pauli component error rates across the device. Finally, Extended Data Fig. S13(h) fits across odd distances to extract the standard per-cycle logical error rate (LER) suppression factor for the Z basis, X basis, and combined mean. We emphasize that shown here is the standard LER suppression factor characterizing the per-cycle rate , whereas in the main text is the block suppression factor evaluated from the optimal memory error probability at the spacetime-optimal cycle count (which includes data qubit state-preparation and measurement (SPAM) errors).
Figure S12. Individual subpatch and basis performance for the stability experiment. (a–e) Logical error probability versus cycles across individual subpatches for (a) (7 grids), (b) (4 grids), (c) (2 grids), (d) (1 grid), and (e) (1 grid). Colors identify subpatches; solid curves and filled circles denote , and dashed curves and open triangles denote ( and are only). (f) Subpatch-averaged decay for (circles, solid fits), (open triangles, dashed fits), and combined mean (dotted fits), fitted to . (g) Block error across subpatches for (blue circles) and (red triangles), with horizontal bars for the , , and combined means. (h) Fast (, top) and slow (, bottom) suppression factors for individual subpatches (faint) and the , , and combined means (1σ error bars).
Figure S13. Individual subpatch performance, basis comparison, and standard LER suppression for surface code memory. (a–e) Logical fidelity versus cycles on a logarithmic axis across subpatches for (a) , (b) , (c) , (d) , and (e) . Colors match Extended Data Fig. S12, with filled circles and solid fits for , and open triangles and dashed fits for . (f) Subpatch-averaged logical fidelity for (circles, solid fits), (open triangles, dashed fits), and combined mean (dotted fits). (g) Per-cycle logical error rate across subpatches for (blue circles) and (red triangles), with , , and combined mean bars. (h) Standard per-cycle LER suppression factor across for , , and combined mean (contrasted with the main-text block at ).
References
- [1]A. Y. Kitaev, Fault-tolerant quantum computation by anyons, Annals of Physics 303, 2 (2003), arXiv:quant-ph/9707021 [quant-ph].
- [2]A. G. Fowler, M. Mariantoni, J. M. Martinis, and A. N. Cleland, Surface codes: Towards practical large-scale quantum computation, Phys. Rev. A 86, 032324 (2012).DOI
- [3]B. M. Terhal, Quantum error correction for quantum memories, Rev. Mod. Phys. 87, 307 (2015).DOI
- [4]Google Quantum AI and Collaborators, R. Acharya, D. A. Abanin, L. Aghababaie-Beni, I. Aleiner, T. I. Andersen, M. Ansmann, F. Arute, K. Arya, A. Asfaw, N. Astrakhantsev, J. Atalaya, R. Babush, D. Bacon, J. C. Bardin, J. Bausch, A. Bengtsson, A. Bilmes, S. Blackwell, S. Boixo, G. Bortoli, A. Bourassa, J. Bovaird, L. Brill, M. Broughton, D. A. Browne, B. Buchea, B. B. Buckley, D. A. Buell, T. Burger, B. Burkett, N. Bushnell, A. Cabrera, J. Campero, H.-S. Chang, Y. Chen, Z. Chen, B. Chiaro, D. Chik, C. Chou, J. Claes, A. Y. Cleland, J. Cogan, R. Collins, P. Conner, W. Courtney, A. L. Crook, B. Curtin, S. Das, A. Davies, L. De Lorenzo, D. M. Debroy, S. Demura, M. Devoret, A. Di Paolo, P. Donohoe, I. Drozdov, A. Dunsworth, C. Earle, T. Edlich, A. Eickbusch, A. M. Elbag, M. Elzouka, C. Erickson, L. Faoro, E. Farhi, V. S. Ferreira, L. Flores Burgos, E. Forati, A. G. Fowler, B. Foxen, S. Ganjam, G. Garcia, R. Gasca, E. Genois, W. Giang, C. Gidney, D. Gilboa, R. Gosula, A. Grajales Dau, D. Graumann, A. Greene, J. A. Gross, S. Habegger, J. Hall, M. C. Hamilton, M. Hansen, M. P. Harrigan, S. D. Harrington, F. J. H. Heras, S. Heslin, P. Heu, O. Higgott, G. Hill, J. Hilton, G. Holland, S. Hong, H.-Y. Huang, A. Huff, W. J. Huggins, L. B. Ioffe, S. V. Isakov, J. Iveland, E. Jeffrey, Z. Jiang, C. Jones, S. Jordan, C. Joshi, P. Juhas, D. Kafri, H. Kang, A. H. Karamlou, K. Kechedzhi, J. Kelly, T. Khaire, T. Khattar, M. Khezri, S. Kim, P. V. Klimov, A. R. Klots, B. Kobrin, P. Kohli, A. N. Korotkov, F. Kostritsa, R. Kothari, B. Kozlovskii, J. M. Kreikebaum, V. D. Kurilovich, N. Lacroix, D. Landhuis, T. Lange-Dei, B. W. Langley, P. Laptev, K.-M. Lau, L. Le Guevel, J. Ledford, J. Lee, K. Lee, Y. D. Lensky, S. Leon, B. J. Lester, W. Y. Li, Y. Li, A. T. Lill, W. P. Livingston, A. Locharla, E. Lucero, D. Lundahl, A. Lunt, S. Madhuk, F. D. Malone, A. Maloney, S. Mandra, J. Manyika, L. S. Martin, O. Martin, S. Martin, C. Maxfield, J. R. McClean, M. McEwen, S. Meeks, A. Megrant, X. Mi, K. C. Miao, A. Mieszala, R. Molavi, S. Molina, S. Montazeri, A. Morvan, R. Movassagh, W. Mruzckiewicz, O. Naaman, M. Neeley, C. Neill, A. Nersisyan, H. Neven, M. Newman, J. H. Ng, A. Nguyen, M. Nguyen, C.-H. Ni, M. Y. Niu, T. E. O’Brien, W. D. Oliver, A. Oprecak, K. Ottosson, A. Petukhov, A. Pizzuto, J. Platt, R. Potter, O. Pritchard, L. P. Pryadko, C. Quintana, G. Ramachandran, M. J. Reagor, J. Redding, D. M. Rhodes, G. Roberts, E. Rosenberg, E. Rosenfeld, P. Roushan, N. C. Rubin, N. Saei, D. Sank, K. Sankaragomathi, K. J. Satzinger, H. F. Schurkus, C. Schuster, A. W. Senior, M. J. Shearn, A. Shorter, N. Shutty, V. Shvarts, S. Singh, V. Sivak, J. Skrzuny, S. Small, V. Smelyanskiy, W. C. Smith, R. D. Somma, S. Springer, G. Sterling, D. Strain, J. Suchard, A. Szasz, A. Sztein, D. Thor, A. Torres, M. M. Torunbalci, A. Vaishnav, J. Vargas, S. Vdovichev, G. Vidal, B. Villalonga, C. Vollgraff Heidweiller, S. Waltman, S. X. Wang, B. Ware, K. Weber, T. Weidel, T. White, K. Wong, B. W. K. Woo, C. Xing, Z. J. Yao, P. Yeh, B. Ying, J. Yoo, N. Yosri, G. Young, A. Zalcmam, Y. Zhang, N. Zhu, and N. Zobrist, Quantum error correction below the surface code threshold, Nature 638, 920 (2025), arXiv:2408.13687 [quant-ph].
- [5]C. Gidney, Stability experiments: The overlooked dual of memory experiments, Quantum 6, 786 (2022).DOI
- [6]P. W. Shor, Scheme for reducing decoherence in quantum computer memory, Phys. Rev. A 52, R2493 (1995).pubmed.ncbi.nlm.nih.gov/9912632
- [7]S. Krinner, N. Lacroix, A. Remm, A. Di Paolo, É. Genois, C. Leroux, C. Hellings, S. Lazar, F. Swiadek, J. Herrmann, G. J. Norris, C. K. Andersen, M. Müller, A. Blais, C. Eichler, and A. Wallraff, Realizing repeated quantum error correction in a distance-three surface code, Nature 605, 669 (2022), arXiv:2112.03708 [quant-ph].DOI
- [8]Google Quantum AI, R. Acharya, I. Aleiner, R. Allen, T. I. Andersen, M. Ansmann, F. Arute, K. Arya, A. Asfaw, J. Atalaya, R. Babbush, D. Bacon, J. C. Bardin, J. Basso, A. Bengtsson, S. Boixo, G. Bortoli, A. Bourassa, J. Bovaird, L. Brill, M. Broughton, B. B. Buckley, D. A. Buell, T. Burger, B. Burkett, N. Bushnell, Y. Chen, Z. Chen, B. Chiaro, J. Cogan, R. Collins, P. Conner, W. Courtney, A. L. Crook, B. Curtin, D. M. Debroy, A. Del Toro Barba, S. Demura, A. Dunsworth, D. Eppens, C. Erickson, L. Faoro, E. Farhi, R. Fatemi, L. Flores Burgos, E. Forati, A. G. Fowler, B. Foxen, W. Giang, C. Gidney, D. Gilboa, M. Giustina, A. Grajales Dau, J. A. Gross, S. Habegger, M. C. Hamilton, M. P. Harrigan, S. D. Harrington, O. Higgott, J. Hilton, M. Hoffmann, S. Hong, T. Huang, A. Huff, W. J. Huggins, L. B. Ioffe, S. V. Isakov, J. Iveland, E. Jeffrey, Z. Jiang, C. Jones, P. Juhas, D. Kafri, K. Kechedzhi, J. Kelly, T. Khattar, M. Khezri, M. Kieferová, S. Kim, A. Kitaev, P. V. Klimov, A. R. Klots, A. N. Korotkov, F. Kostritsa, J. M. Kreikebaum, D. Landhuis, P. Laptev, K.-M. Lau, L. Laws, J. Lee, K. Lee, B. J. Lester, A. Lill, W. Liu, A. Lochala, E. Lucero, F. D. Malone, J. Marshall, O. Martin, J. R. McClean, T. McCourt, M. McEwen, A. Megrant, B. Meurer Acosta, X. Mi, K. C. Miao, M. Mohseni, S. Montazeri, A. Morvan, E. Mount, W. Mruckczkiewicz, O. Naaman, M. Neeley, C. Neill, A. Nersisyan, H. Neven, M. Newman, J. H. Ng, A. Nguyen, M. Nguyen, M. Y. Niu, T. E. O’Brien, A. Opreamcak, J. Platt, A. Petukhov, R. Potter, L. P. Pryadko, C. Quintana, P. Roushan, N. C. Rubin, N. Saei, D. Sank, K. Sankaragomathi, K. J. Satzinger, H. F. Schurkus, C. Schuster, M. J. Shearn, A. Shorter, V. Shvarts, J. Skruzny, V. Smelyanskiy, W. C. Smith, G. Sterling, D. Strain, M. Szalay, A. Torres, G. Vidal, B. Villalonga, C. Vollgraff Heidweiller, T. White, C. Xing, Z. J. Yao, P. Yeh, J. Yoo, G. Young, A. Zalcman, Y. Zhang, and N. Zhu, Suppressing quantum errors by scaling a surface code logical qubit, Nature 614, 676 (2023), arXiv:2207.06431 [quant-ph].
- [9]K. C. Miao, M. McEwen, J. Atalaya, D. Kafri, L. P. Pryadko, A. Bengtsson, A. Opreemcak, K. J. Satzinger, Z. Chen, P. V. Klimov, C. Quintana, R. Acharya, K. Anderson, M. Ansmann, F. Arute, K. Arya, A. Asfaw, J. C. Bardin, A. Bourassa, J. Bovaird, L. Brill, B. B. Buckley, D. A. Buell, T. Burger, B. Burkett, N. Bushnell, J. Campero, B. Chiaro, R. Collins, P. Conner, A. L. Crook, B. Curtin, D. M. Debroy, S. Demura, A. Dunsworth, C. Erickson, R. Fatemi, V. S. Ferreira, L. Flores Burgos, E. Forati, A. G. Fowler, B. Foxen, G. Garcia, W. Giang, C. Gidney, M. Giustina, R. Gosula, A. Grajales Dau, J. A. Gross, M. C. Hamilton, S. D. Harrington, P. Heu, J. Hilton, M. R. Hoffmann, S. Hong, T. Huang, A. Huff, J. Iveland, E. Jeffrey, Z. Jiang, C. Jones, J. Kelly, S. Kim, F. Kostritsa, J. M. Kreikebaum, D. Landhuis, P. Laptev, L. Laws, K. Lee, B. J. Lester, A. T. Lill, W. Liu, A. Lochala, E. Lucero, S. Martin, A. Megrant, X. Mi, S. Montazeri, A. Morvan, O. Naaman, M. Neeley, N. Nersisyan, M. Newman, J. H. Ng, A. Nguyen, M. Nguyen, R. Potter, C. Rocque, P. Roushan, K. Sankaragomathi, H. F. Schurkus, C. Schuster, M. J. Shearn, A. Shorter, N. Shutty, V. Shvarts, J. Skruzny, W. C. Smith, G. Sterling, M. Szalay, D. Thor, A. Torres, T. White, B. W. K. Woo, Z. J. Yao, P. Yeh, J. Yoo, G. Young, A. Zalcman, N. Zhu, N. Zobrist, H. Neven, V. Smelyanskiy, A. Petukhov, A. N. Korotkov, D. Sank, and Y. Chen, Overcoming leakage in quantum error correction, Nature Physics 19, 1780 (2023), arXiv:2211.04728 [quant-ph].DOI
- [10]A. Eickbusch, M. McEwen, V. S. Sivak, A. Bourassa, J. Atalaya, J. Claes, D. Kafri, C. Gidney, C. W. Warren, J. Gross, A. Opreemcak, N. Zobrist, K. C. Miao, G. Roberts, K. J. Satzinger, A. Bengtsson, M. Neeley, W. P. Livingston, A. Greene, R. Acharya, L. Aghababaie Beni, G. Aigeldinger, R. Alcaraz, T. I. Andersen, M. Ansmann, F. Arute, K. Arya, A. Asfaw, R. Babbush, B. Ballard, J. C. Bardin, A. Bilmes, J. Bovaird, D. Bowers, L. Brill, M. Broughton, D. A. Browne, B. Buchea, B. B. Buckley, T. Burger, B. Burkett, N. Bushnell, A. Cabrera, J. Campero, H.-S. Chang, C. Chiaro, L.-Y. Chih, A. Y. Cleland, J. Cogan, R. Collins, P. Conner, W. Courtney, A. L. Crook, B. Curtin, S. Das, A. Del Toro Barba, S. Demura, L. De Lorenzo, A. Di Paolo, P. Donohoe, I. K. Drozdov, A. Dunsworth, A. M. Elbag, M. Elzouka, C. Erickson, V. S. Ferreira, L. Flores Burgos, E. Forati, A. G. Fowler, B. Foxen, S. Ganjam, G. Garcia, R. Gasca, É. Genois, W. Giang, D. Gilboa, R. Gosula, A. Grajales Dau, D. Graumann, T. Ha, S. Habegger, M. C. Hamilton, M. Hansen, M. P. Harrigan, S. D. Harrington, S. Heslin, P. Heu, O. Higgott, R. Hiltermann, J. Hilton, H.-Y. Huang, A. Huff, W. J. Huggins, E. Jeffrey, Z. Jiang, X. Jin, C. Jones, C. Joshi, P. Juhas, A. Kabel, H. Kang, A. H. Karamlou, K. Kechedzhi, T. Khaire, K. Khattar, M. Khezri, S. Kim, B. Kobrin, A. N. Korotkov, F. Kostritsa, J. M. Kreikebaum, V. D. Kurilovich, D. Landhuis, T. Lange-Dei, B. W. Langley, K.-M. Lau, J. Ledford, K. Lee, B. J. Lester, L. Le Guevel, W. Y. Li, A. T. Lill, A. Lochala, E. Lucero, D. Lundahl, A. Lunt, S. Madhuk, A. Maloney, S. Mandrà, L. S. Martin, O. Martin, C. Maxfield, J. R. McClean, S. Meeks, A. Megrant, R. Molavi, S. Molina, S. Montazeri, R. Movassagh, M. Newman, A. Nguyen, M. Nguyen, M. Nguyẹn, M. Nguyan, C.-H. Ni, L. Oas, R. Orosco, K. Ottosson, A. Pizzuto, R. Potter, O. Pritchard, C. Quintana, G. Ramachandran, M. J. Reagor, D. M. Rhodes, E. Rosenberg, E. Rossi, K. Sankaragomathi, H. F. Schurkus, M. J. Shearn, A. Shorter, N. Shutty, V. Shvarts, S. Small, W. C. Smith, S. Springer, G. Sterling, J. Suchard, A. Szasz, A. Sztein, D. Thor, E. Tomita, A. Torres, M. M. Torunbalci, A. Vaishnav, J. Vargas, S. Vdovichev, G. Vidal, C. Vollgraff Heidweiller, S. Waltman, J. Waltz, S. X. Wang, B. Ware, T. Weidel, T. White, K. Wong, B. W. K. Woo, M. Woodson, C. Xing, Z. J. Yao, P. Yeh, B. Ying, J. Yoo, N. Yosri, G. Young, A. Zalcman, Y. Zhang, N. Zhu, S. Boixo, J. Kelly, V. Smelyanskiy, H. Neven, D. Bacon, Y. K. K. Kim, P. Roushan, C. Neill, Y. Chen, and A. Morvan, Demonstration of dynamic surface codes, Nature Physics 21, 1994 (2025), arXiv:2412.14360 [quant-ph].
- [11]T. He, W. Lin, R. Wang, Y. Li, J. Bei, J. Cai, S. Cao, D. Chen, K. Chen, X. Chen, Z. Chen, Z. Chen, Z. Chen, W. Chu, H. Deng, X. Ding, Z. Ding, B. Fan, D. Fan, Y. Fu, D. Gao, M. Gong, J. Gui, C. Guo, S. Guo, L. Han, L. Hong, Y. Hu, H.-L. Huang, Y.-H. Huo, C. Jiang, L. Jiang, T. Jiang, Z. Jiang, H. Jin, D. Li, D. Li, J. Li, J. Li, J. Li, J. Li, N. Li, S. Li, Y. Li, F. Liang, N. Liao, J. Lin, K. Liu, M. Liu, Y. Liu, H. Lou, Y. Ma, K. Nan, M. Nie, L. Niu, W. Peng, H. Qian, H. Rong, T. Rong, H. Shen, Q. Shen, H. Su, F. Su, C. Sun, L. Sun, T. Sun, Y. Sun, Y. Tan, J. Tan, W. Tu, J. Wang, B. Wang, C. Wang, C. Wang, C. Wang, J. Wang, S. Wang, X. Wang, Z. Wei, D. Wu, G. Wu, Y. Wu, Y. Xu, C. Xue, K. Yan, X. Yan, W. Yang, X. Yang, Y. Yang, Y. Ye, Z. Ye, Z. Yi, C. Ying, J. Yu, Q. Yu, X. Zeng, C. Zha, S. Zhan, H. Zhang, H. Zhang, K. Zhang, W. Zhang, Y. Zhang, Y. Zhang, Z. Zhang, G. Zhao, X. Zhao, Y. Zhao, Z. Zhao, L. Zheng, F. Zhou, L. Zhou, N. Zhou, N. Zhou, C. Zhu, Q. Zhu, G. Zou, H. Zou, Q. Zhang, C.-Y. Lu, C.-Z. Peng, F. Chen, X. Zhu, and J.-W. Pan, Experimental Quantum Error Correction below the Surface Code Threshold via All-Microwave Leakage Suppression, Physical Review Letters 135, 260601 (2025).DOI
- [12]J. A. Muniz, D. Crow, H. Kim, J. M. Kindem, W. B. Cairncross, A. Ryou, T. C. Bohdanowicz, C.-A. Chen, Y. Ji, A. M. W. Jones, E. Megidish, C. Nishiguchi, M. Urbanek, L. Wadleigh, T. Wilkason, D. Aasen, K. Barnes, J. M. Bello-Rivas, I. Bloomfield, G. Booth, A. Brown, M. O. Brown, K. Cassella, G. Cowan, J. Epstein, M. Feldkamp, C. Griger, Y. Hassan, A. Heinz, E. Halperin, T. Hofler, F. Hummel, M. Jaffe, E. Kapit, K. Kotru, J. Lauigan, J. Marjanovic, M. Meredith, M. McDonald, R. Morshead, S. Narayanaswami, K. A. Pawlak, K. L. Pudenz, D. Rodríguez Pérez, P. Sabharwal, J. Simon, A. Smull, M. Sorensen, D. T. Stack, M. Stone, L. Taneja, R. J. M. van de Weerdonk, Z. Vendeiro, R. T. Weverka, K. White, T.-Y. Wu, X. Xie, E. Zalys-Geller, X. Zhang, J. K. Bloom, and M. A. Norcia, Repeated ancilla reuse for logical computation on a neutral atom quantum computer, Physical Review X 15, 041040 (2025), arXiv:2506.09936 [quant-ph].arxiv.org/abs/2506.09936
- [13]Atom Computing and Collaborators, Quantum error correction with the toric code (2026), arXiv:2606.04079 [quant-ph].arxiv.org/abs/2606.04079
- [14]L. Caune, L. Skoric, N. S. Blunt, A. Ruban, J. McDaniel, J. A. Valery, A. D. Patterson, A. V. Gramolin, J. Majaniemi, K. M. Barnes, T. Bialas, O. Buğdaycı, O. Crawford, G. P. Gehér, H. Krovi, E. Matekole, C. Topal, S. Poletto, M. Bryant, K. Snyder, N. I. Gillespie, G. Jones, K. Johar, E. T. Campbell, and A. D. Hill, Demonstrating real-time and low-latency quantum error correction with superconducting qubits, Nature Communications 17, 7383 (2026), arXiv:2410.05202 [quant-ph].arxiv.org/abs/2410.05202
- [15]J. M. Drouet, X. C. Kolesnikow, C. K. McLauchlan, G. M. Nixon, S.-H. Lee, D. J. Williamson, S. D. Bartlett, B. J. Brown, and R. Harper, Error correction on an array of superconducting qubits with defective components (2026), arXiv:2607.12118 [quant-ph].arxiv.org/abs/2607.12118
- [16]Y. Xin, S. L. M. van der Meer, M. Serra-Peralta, T. H. F. Vroomans, M. Finkel, H. M. Veen, M. W. Beekman, and L. DiCarlo, Improved Error Correction with Leakage Reduction Units Built into Qubit Measurement in a Superconducting Quantum Processor, PRX Quantum 7, 033062 (2026), arXiv:2511.17460 [quant-ph].DOI
- [17]B. Zhang, G. Liu, G. Bornet, S. P. Horvath, P. Peng, S. Ma, S. Huang, S. Puri, and J. D. Thompson, Logical qubits with erasure conversion using metastable neutral atoms, Nature Physics 22, 910 (2026), arXiv:2506.13724 [quant-ph].DOI
- [18]L. Postler, S. Heußen, I. Pogorelov, M. Rispler, T. Feldker, M. M. Střch, C. D. Marciniak, R. Stricker, M. Ringbauer, R. Blatt, P. Schindler, M. Müller, and T. Monz, Demonstration of fault-tolerant universal quantum gate operations, Nature 605, 675 (2022), arXiv:2111.12654 [quant-ph].DOI
- [19]D. Bluvstein, S. J. Evered, A. A. Geim, S. H. Li, H. Zhou, T. Manovitz, S. Ebadi, M. Cain, M. Kalinowski, D. Hangleiter, J. P. Bonilla Ataides, N. Maskara, I. Cong, X. Gao, P. Sales Rodriguez, T. Karolyshyn, G. Semeghini, M. J. Gullans, M. Greiner, V. Vuletić, and M. D. Lukin, A logical quantum processor based on reconfigurable atom arrays, Nature 626, 58 (2024), arXiv:2312.03818 [quant-ph].
- [20]B. W. Reichardt, A. Paetznick, D. Aasen, I. Basov, J. M. Bello-Rivas, P. Bonderson, R. Chao, W. van Dam, M. B. Hastings, R. V. Mishmash, A. Paz, M. P. da Silva, A. Sundaram, K. M. Svore, A. Vaschillo, Z. Wang, M. Zanner, W. B. Cairncross, C.-A. Chen, D. Crow, H. Kim, J. M. Kindem, J. King, M. McDonald, M. A. Norcia, A. Ryou, M. Stone, L. Wadleigh, K. Barnes, P. Battaglino, T. C. Bohdanowicz, G. Booth, A. Brown, M. O. Brown, K. Cassella, R. Coxe, J. M. Epstein, M. Feldkamp, C. Griger, E. Halperin, A. Heinz, F. Hummel, M. Jaffe, A. M. W. Jones, E. Kapit, K. Kotru, J. Lauigan, M. Li, J. Marjanovic, E. Megidish, M. Meredith, R. Morshead, J. A. Muniz, S. Narayanaswami, C. Nishiguchi, T. Paule, K. A. Pawlak, K. L. Pudenz, D. Rodríguez Pérez, J. Simon, A. Smull, D. Stack, M. Urbanek, R. J. M. van de Weerdonk, Z. Vendeiro, R. T. Weverka, T. Wilkason, T.-Y. Wu, X. Xie, E. Zalys-Geller, X. Zhang, and B. J. Bloom, Fault-tolerant quantum computation with a neutral atom processor (2024), arXiv:2411.11822 [quant-ph].arxiv.org/abs/2411.11822
- [21]C. Ryan-Anderson, N. C. Brown, C. H. Baldwin, J. M. Dreiling, C. Foltz, J. P. Gaebler, T. M. Gatterman, N. Hewitt, C. Holliman, C. V. Horst, J. Johansen, D. Lucchetti, T. Mengle, M. Matheny, Y. Matsuoka, K. Mayer, M. Mills, S. A. Moses, B. Neyenhuis, J. Pino, P. Siegfried, R. P. Stutz, J. Walker, and D. Hayes, High-fidelity teleportation of a logical qubit using transversal gates and lattice surgery, Science 385, 1327 (2024), arXiv:2404.16728 [quant-ph].
- [22]L. Daguerre, R. Blume-Kohout, N. C. Brown, D. Hayes, and I. H. Kim, Experimental Demonstration of High-Fidelity Logical Magic States from Code Switching, Physical Review X 15, 041008 (2025), arXiv:2506.14169 [quant-ph].DOI
- [23]N. Lacroix, A. Bourassa, F. J. H. Heras, L. M. Zhang, J. B. Schuch, A. W. Senyshyn, T. Edlich, N. Shuty, V. S. Vask, A. Bengtsson, M. McEwen, O. Higgott, D. Kafri, J. Claes, A. Morvan, Z. Chen, A. Zalcman, S. Madhuk, R. Acharya, L. Aghababaie Beni, G. Ajeldinger, R. Alcaraz, T. I. Andersen, M. Ansmann, F. Arute, K. Arya, A. Asfaw, J. Atalaya, R. Babbush, B. Ballard, J. C. Bardin, A. Bilmes, S. Blackwell, J. Bovaird, D. Bowers, L. Brill, M. Broughton, D. A. Browne, B. Buchea, B. B. Buckley, T. Burger, B. Burkett, N. Bushnell, A. Cabrera, J. Campero, H.-S. Chang, B. Chiaro, L.-Y. Chih, A. Y. Cleland, J. Cogan, R. Collins, P. Conner, W. Courtney, A. L. Crook, B. Curtin, S. Das, S. Demura, L. De Lorenzo, A. Di Paolo, P. Donohoe, I. Drozdov, A. Dunsworth, A. Fickbusch, A. M. Elbag, M. Elbag, M. Erickson, V. S. Ferreira, L. Flores Burgos, E. Forati, A. G. Fowler, B. Foxen, S. Ganjam, G. Garcia, R. Gasca, É. Genois, W. Giang, D. Gilboa, R. Gosula, A. Grajales Dau, D. Graumann, A. Greene, J. A. Gross, T. Ha, S. Habe gger, M. Hansen, M. P. Harrigan, S. D. Harrington, S. Heslin, P. Heu, R. Hiltermann, J. Hilton, S. Hong, H.-Y. Huang, A. Huff, W. J. Huggins, E. Jeffrey, Z. Jiang, X. Jin, C. Joshi, P. Juhas, A. Kabel, H. Kang, A. H. Karamlou, K. Kechedzhi, T. Khaire, T. Khattar, M. Khezri, S. Kim, P. V. Klimov, B. Kobrin, A. N. Korotkov, F. Kostritsa, J. M. Kreikebaum, V. D. Kurilovich, D. Landhuis, T. Lange-Dei, B. W. Langley, P. Laptev, K.-M. Lau, J. Ledford, K. Lee, B. J. Lester, L. Le Guevel, W. Y. Li, Y. Li, A. T. Lill, W. P. Livingston, A. Locharla, E. Lucero, D. Lundahl, A. Lunt, A. Maloney, S. Mandrà, L. S. Martin, O. Martin, C. Maxfield, J. R. McClean, S. Meeks, A. Megrant, K. C. Miao, R. Molavi, S. Molina, S. Montazeri, R. Movassagh, C. Neill, M. Newman, A. Nguyen, M. Nguyen, C.-H. Ni, M. Y. Niu, L. Oas, W. D. Oliver, R. Orosco, K. Ottosson, A. Pizzuto, R. Potter, O. Pritchard, C. Quintana, G. Ramachandran, M. J. Reagor, R. Resnick, D. M. Rhodes, G. Roberts, E. Rosenberg, E. Rosenfeld, E. Rossi, P. Roushan, K. Sankaragomathi, H. F. Schurkus, M. J. Shearn, A. Shorter, V. Shvarts, S. Small, W. C. Smith, S. Springer, G. Sterling, J. Suchard, A. Szasz, A. Sztein, D. Thor, E. Tomita, A. Torres, M. M. Torunbalci, A. Vaishnav, J. Vargas, S. Vdovichev, G. Vidal, C. Vollgraff Heidweiller, S. Waltman, J. Waltz, X. Wang, B. Ware, T. Weidel, T. White, K. Wong, B. W. K. Woo, M. Woodson, C. Xing, Z. J. Yao, P. Yeh, B. Ying, J. Yoo, N. Yosri, G. Young, Y. Zhang, N. Zhu, N. Zobrist, H. Neven, P. Kohli, A. Davies, S. Boixo, J. Kelly, C. Jones, C. Gidney, and K. J. Satzinger, Scaling and logic in the colour code on a superconducting quantum processor, Nature 645, 614 (2025).DOI
- [24]E. Rosenfeld, C. Gidney, G. Roberts, A. Morovan, N. Lacroix, D. Kafri, J. Marshall, M. Li, V. Sivak, D. Abanin, A. Abbas, R. Acharya, L. Aghababaie Beni, G. Aigeldinger, R. Alcaraz, S. Alcaraz, T. I. Andersen, M. Ansmann, F. Arute, K. Arya, W. Askew, N. Astrakhantsev, J. Atalaya, R. Babbush, B. Ballard, J. C. Bardin, H. Bates, A. Bengtsson, M. B. Karimi, A. Bilmes, S. Bilodeau, F. Borjans, J. Bovoird, D. Bowers, L. Brill, P. Brooks, M. Broughton, D. A. Browne, B. Buchea, B. B. Buckley, T. Burger, B. Burkett, N. Bushnell, J. Busnaina, A. Cabrera, J. Campero, H.-S. Chang, S. Chen, Z. Chen, B. Chiaro, L.-Y. Chih, A. Y. Cleland, B. Cochrane, M. Cockrell, J. Cogan, P. Conner, H. Cook, R. G. Cortiñas, W. Courtney, A. L. Crook, B. Curtin, M. Damyanov, S. Das, D. M. Debroy, S. Demura, P. Donohoe, I. Drozdov, A. Dunsworth, V. Ehimhen, A. Eickbusch, A. M. Elbag, L. Ella, M. Elzouka, D. Enriquez, C. Erickson, L. Faoro, V. S. Ferreira, M. Flores, L. Flores Burgos, S. Fontes, E. Forati, J. Ford, B. Foxen, M. Fukami, A. W. L. Fung, L. Fuste, S. Ganjam, G. Garcia, C. Garrick, R. Gasca, H. Gehring, R. Geiger, E. Genois, W. Giang, D. Gilboa, J. E. Goedder, E. C. Gonzales, R. Gosula, S. J. de Graaf, A. Grajales Dau, D. Graumann, J. Grebel, A. Greene, J. A. Gross, J. Guerrero, L. Le Guevel, T. Ha, S. Habe gger, T. Hadick, A. Hadjikhani, M. C. Hamilton, M. Hansen, M. P. Harrigan, S. D. Harrington, J. Hartshorn, S. Heslin, P. Heu, O. Higgott, R. Hiltermann, J. Hilton, H.-Y. Huang, M. Hucka, C. Hudspeth, A. Huff, W. J. Huggins, L. B. Ioffe, E. Jeffrey, S. Jevons, Z. Jiang, X. Jin, C. Joshi, P. Juhas, A. Kabel, H. Kang, K. Kang, A. H. Karamlou, R. Kaufman, K. Kechedzhi, T. Khattar, M. Khezri, S. Kim, P. V. Klimov, C. M. Knaut, B. Kobrin, A. N. Korotkov, F. Kostritsa, J. M. Kreikebaum, R. Kudo, B. Kueffler, A. Kumar, V. D. Kurilovich, W. Kutsko, T. Lange-Dei, B. W. Langley, P. Laptev, K.-M. Lau, E. Leavell, J. Ledford, J. Lee, K. Lee, B. J. Lester, W. Leung, L. Li, W. Y. Li, A. T. Lill, W. P. Livingston, M. T. Lloyd, A. Locharla, L. De Lorenzo, E. Lucero, D. Lundahl, A. Lunt, S. Madhuk, A. Maiti, A. Maloney, S. Mandrà, L. S. Martin, O. Martin, E. Mascot, P. Masih Das, D. Maslov, M. Mathews, C. Maxfield, J. R. McClean, M. McEwen, S. Meeks, A. Megrant, K. C. Miao, Z. K. Minev, R. Molavi, S. Molina, S. Montazeri, C. Neill, M. Newman, A. Nguyen, M. Nguyen, C.-H. Ni, M. Y. Niu, N. Noll, L. Oas, W. D. Oliver, R. Orosco, K. Ottosson, A. Pagano, A. Di Paolo, S. Peek, D. Peterson, A. Pizzuto, E. Portoles, R. Potter, O. Pritchard, M. Qian, C. Quintana, G. Ramachandran, A. Ranadive, M. J. Reagor, R. Resnick, D. M. Rhodes, D. Riley, R. Rodriguez, E. Ropes, L. B. De Rose, E. Rosenberg, D. Rosenstock, E. Rossi, P. Roushan, D. A. Rower, R. Salazar, K. Sankaragomathi, M. C. Sarihah, M. Schaefer, S. Schroeder, H. F. Schurkus, A. Shahingohar, M. J. Shearn, A. Shorter, N. Shutty, V. Shvarts, S. Small, W. C. Smith, D. A. Sobel, B. Spells, S. Springer, G. Sterling, J. Suchard, A. Szasz, A. Sztein, M. Taylor, J. P. Thiruman, D. Thor, D. Timucin, E. Tomita, A. Torres, M. M. Torunbalci, H. Tran, A. Vaishnav, J. Vargas, S. Vdovichev, G. Vidal, B. Villalonga, C. Vollgraff Heidweiller, M. Voorhees, S. Waltman, J. Waltz, S. X. Wang, D. Wang, B. Ware, J. D. Watson, Y. Wei, T. Weidel, T. White, K. Wong, B. W. K. Woo, C. J. Wood, M. Woodson, C. Xing, Z. J. Yao, P. Yeh, B. Ying, J. Yoo, N. Yosri, E. Young, G. Young, A. Zalcman, R. Zhang, Y. Zhang, N. Zhu, N. Zobrist, Z. Zou, H. Neven, S. Boixo, C. Jones, J. Kelly, A. Bourassa, and K. J. Satzinger, Magic state cultivation on a superconducting quantum processor, Science 393, 71 (2026), arXiv:2512.13908 [quant-ph].arxiv.org/abs/2512.13908
- [25]P. Sales Rodriguez, J. M. Robinson, P. N. Jepsen, Z. He, C. Duckering, C. Zhao, K.-H. Wu, J. Campo, K. Bagnall, M. Kwon, T. Karolyshyn, P. Weinberg, M. Cain, S. J. Evered, A. A. Geim, M. Kalinowski, S. H. Li, T. Manovitz, J. Amato-Grill, J. I. Basham, L. Bernstein, B. Braverman, A. Bylinskii, A. Choukri, R. J. DeAngelo, F. Fang, C. Fieweger, P. Frederick, D. Haines, M. Hamdan, J. Hammett, N. Hsu, M.-G. Hu, F. Huber, N. Jia, D. Kedar, M. Kornjaca, F. Liu, J. Long, J. Lopatin, P. L. S. Lopes, X.-Z. Luo, T. Macrì, O. Marković, L. A. Martínez-Martínez, X. Meng, S. Ostermann, E. Ostroumov, D. Paquette, Z. Qiang, V. Shofman, A. Singh, M. Singh, N. Sinha, H. Thoreen, N. Wan, Y. Wang, D. Waxman-Len, T. Wong, J. Wurtz, A. Zhdanov, L. Zheng, M. Greiner, A. Keesling, N. Gemelke, V. Veluti, T. Kitagawa, S.-T. Wang, D. Bluvstein, M. D. Lukin, A. Lukin, H. Zhou, and S. H. Cantú, Experimental demonstration of logical magic state distillation, Nature 645, 620 (2025), arXiv:2412.15165 [quant-ph].DOI
- [26]H. Zhou, C. Zhao, M. Cain, D. Bluvstein, N. Maskara, C. Duckering, H.-Y. Hu, S.-T. Wang, A. Kubica, and M. D. Lukin, Low-overhead transversal fault tolerance for universal quantum computation, Nature 646, 303 (2025), arXiv:2406.17653 [quant-ph].DOI
- [27]I. Besedin, M. Kerschbaum, J. Knoll, I. Hesner, L. Bödeker, L. Colmenarez, L. Hofele, N. Lacroix, C. Hellings, F. Swiadek, G. J. Norris, A. Remm, S. Krinner, A. Blais, C. Eichler, and A. Wallraff, Lattice surgery realized on two distance-three repetition codes with superconducting qubits, Nature Physics 22, 189 (2026), arXiv:2501.04612 [quant-ph].arxiv.org/abs/2501.04612
- [28]R. Harper, C. Lainé, E. T. Hockings, C. McLauchlan, G. M. Nixon, B. J. Brown, D. Bartlett, Characterising the failure mechanisms of error-corrected quantum logic gates, Nature Communications 17, 5039 (2026), arXiv:2504.07258 [quant-ph].DOI
- [29]W. Lin, S. Guo, Y. Ma, Z. Yi, K. Zhang, J. Bei, J. Cai, S. Cao, D. Chen, G. Chen, J. Chen, K. Chen, X. Chen, Z. Chen, Z. Chen, Z. Chen, W. Chu, H. Deng, X. Ding, Z. Ding, Y. Du, B. Fan, D. Fan, Y. Fu, D. Gao, M. Gong, J. Gui, C. Guo, L. Han, T. He, L. Hong, Y. Hu, H.-L. Huang, Y.-H. Huo, C. Jiang, L. Jiang, T. Jiang, W. Jiang, Z. Jiang, D. Li, D. Li, J. Li, J. Li, J. Li, S. Li, W. Li, X. Li, Y. Li, Y. Li, F. Liang, N. Liao, J. Lin, M. Liu, Y. Liu, H. Lou, K. Nan, M. Nie, L. Niu, W. Peng, H. Qian, H. Rong, T. Rong, Y. Shao, H. Shen, Q. Shen, G. Song, H. Su, F. Su, C. Sun, D. Sun, L. Sun, T. Sun, Y. Sun, Y. Tan, J. Tan, S. Tang, Y. Tang, W. Tu, J. Wang, B. Wang, C. Wang, C. Wang, C. Wang, J. Wang, R. Wang, S. Wang, X. Wang, Z. Wang, Z. Wei, G. Wu, Y. Wu, H. Xia, S. Xia, S. Xie, Z. Xie, L. Xiong, J. Xu, Y. Xu, Y. Xu, C. Xue, K. Yan, X. Yan, W. Yang, X. Yang, Y. Yang, Y. Yang, Y. Yang, Y. Yang, Y. Yang, J. Yu, Q. Yu, C. Zha, S. Zhan, C. Zhang, F. Zhang, H. Zhang, H. Zhang, H. Zhang, K. Zhang, Q. Zhang, S. Zhang, W. Zhang, X. Zhang, Y. Zhang, Y. Zhang, Z. Zhang, G. Zhao, X. Zhao, Y. Zhao, Z. Zhao, L. Zheng, F. Zhou, L. Zhou, N. Zhou, C. Zhu, Q. Zhu, Y. Zhu, G. Zou, H. Zou, Q. Zhang, C.-Y. Lu, J. Chen, C.-Z. Peng, F. Chen, X. Zhu, and J.-W. Pan, Surface code logical operations on a superconducting quantum processor (2026), arXiv:2607.01473 [quant-ph].arxiv.org/abs/2607.01473
- [30]C. K. McLauchlan, G. M. Nixon, J. M. Drouet, X. C. Kolesnikow, S.-H. Lee, J. Raftery, K. Siva, S. D. Bartlett, B. J. Brown, and R. Harper, Error-Corrected Memory and Logic on a Heavy-Hex Superconducting-Qubit Processor (2026), arXiv:2610.11658 [quant-ph].arxiv.org/abs/2610.11658
- [31]M. Paetznick, B. W. Reichardt, M. P. da Silva, C. Ryan-Anderson, D. Aasen, J. M. Bello-Rivas, J. P. Campora, R. Chao, A. Chernoguzov, W. van Dam, J. M. Dreiling, C. Foltz, F. Frölich, J. P. Gaebler, T. M. Gatterman, L. Grans-Samuelsson, D. Gresh, D. Hayes, N. Hewitt, C. A. Holliman, C. V. Horst, J. Johansen, D. Lucchetti, Y. Matsuoka, M. Mills, S. A. Moses, B. Neyenhuis, A. Paz, J. Pino, P. E. Siegfried, A. Sundaram, D. Tom, Z. Wang, S. J. Wernli, M. Zanner, R. P. Stutz, and K. M. Svore, Improved quantum processor logical error rates via correction and detection, Nature 654, 349 (2026), arXiv:2404.02280 [quant-ph].
- [32]M. McEwen, L. Faaro, K. Arya, A. Dunsworth, T. Huang, S. Kim, B. Burkett, A. Fowler, F. Arute, J. C. Bardin, A. Bengtsson, A. Bilmes, B. B. Buckley, N. Bushnell, Z. Chen, R. Collins, S. Demura, A. R. Derk, C. Erickson, M. Giustina, S. D. Harrington, S. Hong, E. Jeffrey, J. Kelly, P. V. Klimov, F. Kostrissa, P. Laptev, A. Locharla, X. Mi, K. C. Miao, S. Montazeri, J. Mutus, O. Naaman, M. Neeley, C. Neill, A. Opremcak, C. Quintana, N. Redd, P. Roushan, D. Sank, K. J. Satzinger, V. Shvarts, T. White, Z. J. Yao, P. Yeh, J. Yoo, Y. Chen, V. Smelyanskiy, J. M. Martinis, H. Neven, A. Megrant, L. Ioffe, and R. Barends, Resolving catastrophic error bursts from cosmic rays in large arrays of superconducting qubits, Nat. Phys. 18, 107 (2022).DOI
- [33]V. D. Kurilovich, G. Roberts, L. S. Martin, M. McEwen, A. Eickbusch, L. Faoro, L. B. Ioffe, J. Atalaya, A. Bilmes, J. M. Kreikebaum, A. Bengtsson, P. Klimov, M. Neeley, W. M. Mueckcziewicz, K. Miao, I. L. Aleiner, J. Kelly, Y. Chen, K. Satzinger, and A. Opremcak, Correlated Phase Error Bursts in a Gap-Engineered Superconducting Qubit Array, Physical Review X 16, 021025 (2026), arXiv:2506.18228 [quant-ph].arxiv.org/abs/2506.18228
- [34]B. Domokos, Á. Márton, and J. K. Asbóth, Characterization of errors in a CNOT between surface code patches, Quantum 8, 1577 (2024), arXiv:2407.03929 [quant-ph].
- [35]M. McEwen, D. Bacon, and C. Gidney, Relaxing Hardware Requirements for Surface Code Circuits using Time-dynamics, Quantum 7, 1172 (2023).DOI
- [36]C. Gidney, M. Newman, P. Brooks, and C. Jones, Yoked surface codes, Nature Communications 16, 4498 (2025), arXiv:2312.04522 [quant-ph].arxiv.org/abs/2312.04522
- [37]Y. Akahoshi, R. Toshio, J. Fujisaki, H. Oshima, S. Sato, and K. Fujii, Runtime reduction in lattice surgery utilizing timelike soft information, Physical Review Research 8, 033085 (2026), arXiv:2510.21149 [quant-ph].arxiv.org/abs/2510.21149
- [38]M. McEwen, D. Kafri, Z. Chen, J. Atalaya, K. J. Satzinger, C. Quintana, P. V. Klimov, D. Sank, C. Gidney, A. G. Fowler, F. Arute, K. Arya, B. Buckley, B. Burkett, N. Bushnell, B. Chiaro, R. Collins, S. Demura, A. Dunsworth, C. Erickson, B. Foxen, M. Giustina, T. Huang, E. Jeffrey, S. Kim, K. Kechedzhi, F. Kostrissa, P. Laptev, A. Megrant, X. Mi, J. Mutus, O. Naaman, M. Neeley, C. Neill, M. Niu, A. Paler, N. Redd, P. Roushan, T. C. White, J. Yao, P. Yeh, A. Zalcman, Y. Chen, V. N. Smelyanskiy, J. M. Martinis, H. Neven, J. Kelly, A. N. Korotkov, A. G. Petukhov, and R. Barends, Removing leakage-induced correlated errors in superconducting quantum error correction, Nature Communications 12, 1761 (2021), arXiv:2102.06131 [quant-ph].DOI
- [39]O. Higgott and C. Gidney, Sparse Blossom: correcting a million errors per core second with minimum-weight matching, Quantum 9, 1600 (2025), arXiv:2303.15933 [quant-ph].DOI
- [40]C. Gidney, M. Newman, A. Fowler, J. M. Broughton, A fault-tolerant honeycomb memory, Quantum 5, 605 (2021).DOI
- [41]Except for w = 6, 7, for which device layout constraints limit the experiment to the Z basis only.
- [42]V. Sivak, A. Morvan, M. Broughton, R. G. Cortiñas, J. Bausch, A. W. Senior, M. Neeley, A. Eickbusch, N. Shutt, L. A. Beni, et al., Reinforcement learning control of quantum error correction, Nature 655, 879 (2026), arXiv:2511.08493 [quant-ph].arxiv.org/abs/2511.08493
- [43]J. Bausch, A. W. Senior, F. J. H. Heras, T. Edlich, A. Davies, S. Blackwell, G. Holland, D. Hassabis, P. Kohli, M. Newman, C. Jones, K. Satzinger, M. Y. Niu, D. Kafri, J. Atalaya, C. Gidney, S. Boixo, and H. Neven, Learning high-accuracy error decoding for quantum processors, Nature 635, 834 (2024).
- [44]M. McEwen, K. C. Miao, J. Atalaya, A. Bilmes, A. Crook, J. Bovaird, J. M. Kreikebaum, N. Zobrist, E. Jeffrey, B. Ying, A. Bengtsson, H.-S. Chang, A. Dunsworth, J. Kelly, Y. Zhang, E. Forati, R. Acharya, J. Iv eland, W. Liu, S. Kim, B. Burkett, A. Megrant, Y. Chen, C. Neill, D. Sank, M. Devoret, and A. Opremcak, Resisting high-energy impact events through gap engineering in superconducting qubit arrays, Phys. Rev. Lett. 133, 240601 (2024).DOI
- [45]H. Ali, J. F. Marques, O. Crawford, J. Majaniemi, M. Serra-Peralta, D. Byfield, B. M. Varbanov, B. M. Terhal, L. DiCarlo, and E. T. Campbell, Reducing the Error Rate of a Superconducting Logical Qubit Using Analog Readout Information, Physical Review Applied 22, 044031 (2024), arXiv:2403.00706 [quant-ph].DOI
- [46]R. Harper, X. C. Kolesnikov, C. K. McLauchlan, G. M. Nixon, S. D. Bartlett, and B. J. Brown, Entangled logical qubits on a superconducting quantum processor (2026), arXiv:2610.32347 [quant-ph].arxiv.org/abs/2610.32347
- [47]C. Gidney, Stim: a fast stabilizer circuit simulator, Quantum 5, 497 (2021).DOI

















