Surface code scaling on heavy-hex superconducting quantum processors Authors: Arian Vezvaee, Cesar Benito, Mario Morford-Oberst, Alejandro Bermudez, Daniel A. Lidar Published: 2026 ABSTRACT Demonstrating subthreshold scaling of a surface-code quantum memory on hardware whose native connectivity does not match the code remains a central challenge. We address this on IBM heavy-hex superconducting processors by co-designing the code embedding and control: a depth-minimizing SWAP-based “fold-unfold” embedding that uses bridge ancillas, together with robust, gap-aware dynamical decoupling (DD). We show that anisotropic scaling from distance 3 to (dx = 3, dz = 5) and (dx = 5, dz = 3) improves protection of Z- and X-basis logical states, respectively, and a calibrated ~ 30% noise reduction would enable isotropic (5, 5)-versus-(3, 3) scaling on this heavy-hex layout. We show that DD suppresses coherent ZZ crosstalk and non-Markovian dephasing during idle gaps, eliminating spurious subthreshold claims. To quantify performance, we derive an entanglement fidelity metric which reveals that widely used single-parameter suppression-factor fits can mischaracterize code performance. Our results provide a path to robust subthreshold surface-code scaling through optimized DD on non-native architectures. Introduction The discovery of quantum error correction (QEC)1–3 enables fault-tolerant quantum computing4–6, which “fights entanglement with entanglement”7 by distributing quantum information among multiple physical qubits, and can safeguard long computations against the accumulation of errors. Numerous experiments have implemented various QEC codes8. Such [[n, k, d]] codes employ n physical qubits to encode k logical qubits and can correct any errors on up to (d − 1)/2 of the physical qubits, where the code distance d9 is odd. Topological codes10, which combine local error-syndrome checks with global information encoding, form a particularly promising class. Perhaps the best known example from this class is the surface code11–13, which has recently been the subject of intense experimental attention14–19. The surface code uses n = d2 data qubits and d2 ancilla qubits to encode k = 1 logical qubits, and leads to a relatively high error threshold20 and simple fault-tolerant implementation21 in comparison to other codes. A more general version of the surface code we use in this work has two different distance parameters, dx and dz for bit and phase-flip errors, respectively, and 2dxdz physical qubits. When the error rate is below the threshold, growing the code distance leads, in theory, to exponential error suppression12. The consequences of subthreshold distance-scaling were demonstrated only very recently by Google using a superconducting quantum processing unit (QPU): the error rate of a logical qubit was shown to break even as d increased from 3 to 515 and, subsequently, to exhibit a persistent twofold improvement from d = 3 to 5 and then from 5 to 717, in line with the expected exponential improvement. Other subthreshold demonstrations of surface and color codes followed, using superconducting qubits22,23 or Rydberg atoms19. A critical aspect of achieving these milestones with Google’s QPUs is the chip’s qubit connectivity graph: a 2D square lattice of transmon qubits designed to precisely mirror the connectivity of the surface code. This allowed for the parallelization of nearly all operations in the QEC cycles, thus minimizing idle gaps (i.e., periods during which a qubit undergoes no gates) and the associated accumulation of errors. Implementing the surface code on other QPUs with different fixed connectivities is a problem of both fundamental interest and practical importance. In light of applications, QPU design may be driven by considerations other than optimal surface code performance, e.g., to avoid frequency crowding affecting crosstalk and gate performance24 and to reduce the density of circuit elements, easing thermal management in lithographic fabrication. For example, IBM’s superconducting QPUs are based on the heavy-hex lattice [illustrated in Fig. 1a], in which transmon qubits are arranged on the sites and links of a honeycomb lattice. An experimental demonstration of subthreshold scaling on heavy-hex devices presents significant challenges. In particular, the reduced connectivity can lead to substantial delays to accommodate state transfer between non-neighboring qubits. The exposure to noise during these additional idle gaps hinders demonstrations of subthreshold scaling18, suggesting that it is crucial to optimize the embedding strategy to minimize the required circuit depth. The lightweight SWAP-based embedding devised in ref. 25 was shown theoretically, on the basis of a multi-parameter, Markovian (stochastic Pauli channels) circuit-level noise model, to outperform other QEC codes better adapted to the heavy-hex architecture, as well as alternative flag-qubit-based embedding protocols. Thus, we implement this embedding here.Fig. 1Surface code on heavy-hex.a Schematic of the surface codes implemented in this work on IBM's Heron-generation QPUs. The red boundary defines the (3, 5) surface code. The yellow boundary shows one of the possible sublattice (3, 3) codes that fits within the larger code. The other two (3, 3) sublattices are not shown. The green boundary defines the (5, 3) surface code. Similarly, there are three (3, 3) sublattices that fit within the (5, 3) code (not shown). b Stabilizer measurement construction for a heavy-hex surface code. The circuit combines the middle-out and traditional ancilla-based syndrome extraction schemes39, adapted to the heavy-hex connectivity25. Next-nearest-neighbor CNOTs are implemented via bridge qubits using SWAP gates, and the resultant circuit is then simplified. Bridge qubits are not measured. Ancilla qubits are never reset; instead, we apply software processing to measurement outcomes depending on previous measurements. The bridge ancilla is not measured; all measurements are on the check ancilla. c N = 0: logical state preparation via two stabilizer rounds, starting from the physical \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\left\vert 0\right\rangle }^{n}$$\end{document}0n state, followed by a measurement of data qubits. The goal of this cycle is SPAM calibration. d N ≥ 1: QEC cycles, each comprising two stabilizer rounds, followed by a measurement of all the data qubits at the end of the N'th cycle. Only half the stabilizers are measured in parallel; a full cycle is two rounds. Surface code on heavy-hex. a Schematic of the surface codes implemented in this work on IBM's Heron-generation QPUs. The red boundary defines the (3, 5) surface code. The yellow boundary shows one of the possible sublattice (3, 3) codes that fits within the larger code. The other two (3, 3) sublattices are not shown. The green boundary defines the (5, 3) surface code. Similarly, there are three (3, 3) sublattices that fit within the (5, 3) code (not shown). b Stabilizer measurement construction for a heavy-hex surface code. The circuit combines the middle-out and traditional ancilla-based syndrome extraction schemes39, adapted to the heavy-hex connectivity25. Next-nearest-neighbor CNOTs are implemented via bridge qubits using SWAP gates, and the resultant circuit is then simplified. Bridge qubits are not measured. Ancilla qubits are never reset; instead, we apply software processing to measurement outcomes depending on previous measurements. The bridge ancilla is not measured; all measurements are on the check ancilla. c N = 0: logical state preparation via two stabilizer rounds, starting from the physical \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\left\vert 0\right\rangle }^{n}$$\end{document}0n state, followed by a measurement of data qubits. The goal of this cycle is SPAM calibration. d N ≥ 1: QEC cycles, each comprising two stabilizer rounds, followed by a measurement of all the data qubits at the end of the N'th cycle. Only half the stabilizers are measured in parallel; a full cycle is two rounds. In this work, we demonstrate subthreshold scaling in the heavy-hex lattice as we independently increase the code’s distance parameters dx or dz for bit- and phase-flip error types. We do not find compelling evidence of basis-independent subthreshold scaling, as the logical states in a basis orthogonal to the direction of growth only suffer from the increased circuit complexity without benefiting from any increased protection. A crucial ingredient in our demonstration is dynamical decoupling (DD)26–29, a technique that is particularly well suited for the suppression of coherent and non-Markovian noise, e.g., due to residual coherent coupling between qubits, or due to a temporally correlated environment with a structured spectral noise density30. We use robust DD variants31,32, which we tailor for our surface code quantum memory, optimizing their integration with the SWAP-based embedding on IBM’s 156-qubit ‘Heron’ QPUs. Although these devices are 6 qubits short of allowing a fully isotropic scaling from d = 3 to d = 5, they do allow for scaling from a 37-qubit embedding of the d = 3 surface code to 65-qubit embeddings of both (dx, dz) = (3, 5) and (dx, dz) = (5, 3) codes. Below threshold, these anisotropic code scalings should endow the encoded logical information with improved protection to either bit- or phase-flip errors, a necessary condition for entering the subthreshold isotropic scaling regime. Our confirmation of this expectation constitutes our main result. We report our quantum memory results for up to 10 full cycles of QEC on ibm_aachen, reaching depths in excess of 140, and a total of 2200 entangling gates. A discussion of similar experiments using other Heron QPUs is given in ref. 33. We extract the logical error probabilities after encoding logical states in the computational, i.e., Z-basis \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert 0\right\rangle }$$\end{document}0¯, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert 1\right\rangle }$$\end{document}1¯) or Hadamard, i.e., X-basis \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert+\right\rangle }$$\end{document}+¯, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert -\right\rangle }$$\end{document}−¯), applying repetitive QEC cycles, and performing a final projective measurement of the data qubits. Note that, in contrast to the X- and Z-basis, the surface-code preparation of Y-basis logical states is significantly more complex, as it cannot be achieved directly within the code’s stabilizer structure and instead requires logical state injection13 or twist defect braiding34,35. The inaccessibility of all six cardinal logical states prevents a complete model-independent estimate of the average channel fidelity36, and other more phenomenological metrics have been proposed. A widely used metric is the suppression factor Λ, calculated as the ratio of logical errors per QEC cycle of two codes of successive odd distances, each averaged over the available X and Z basis states, typically just \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert 0\right\rangle }$$\end{document}0¯ and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert+\right\rangle }$$\end{document}+¯15,17,19,23. We critically assess whether this basis-averaged suppression factor, which is furthermore based on a phenomenological one-parameter fit that assumes stationary (cycle-independent) errors, functions as a reliable and robust performance metric. In the case of IBM QPUs, we show that this is not the case and provide a first-principles derivation of an alternative metric based on the entanglement fidelity (EF)36 that, under fairly general logical error model assumptions (clarified below), allows for a faithful and accurate characterization of our QEC experiments. We note that while the EF was introduced in ref. 36, here we adapt it into a practical benchmarking framework for multi-cycle QEC memory experiments. In particular, we show how EF can be bounded from experimentally accessible logical X- and Z-basis data in a SPAM-aware way, without assuming cycle-independent or unital noise. We use this metric to identify a three-parameter fitting model that separates SPAM from the QEC per-cycle logical error, and use it to demonstrate a reduction in logical error probabilities as we integrate our SWAP-based embedding with robust DD while scaling to the two larger surface codes. In all QPUs used in this work, we identify non-Markovian dephasing and coherent ZZ-crosstalk as the main sources of hardware errors, leading to a dominance of Z-type logical errors. We demonstrate that the incorporation of robust DD in QEC circuits is essential for subthreshold performance. DD suppresses coherent errors, which are underestimated by randomized benchmarking37. In addition, DD suppresses incoherent non-Markovian errors occurring during the extra idle gaps of the SWAP-embedding circuits, which are particularly biased towards single- and two-qubit dephasing. DD suppresses these errors to a level where the surface code can correct the remaining Markovian errors, and to the point where the benefits of anisotropic surface code scaling become more important. We also find that claiming genuine subthreshold scaling requires careful consideration of DD effects. Without DD, logical error rates can decrease with increasing code distance, apparently demonstrating subthreshold behavior. However, when DD is applied, this trend may reverse: although DD reduces errors across all code distances, it can benefit smaller codes disproportionately, eliminating the apparent scaling advantage of larger codes. This reveals the possibility of spurious claims of subthreshold scaling: behavior that appears subthreshold without DD but ceases to be so when DD is properly included. Thus, our results offer insight into the interplay between QEC and error suppression, and the importance of fully optimizing the latter before a claim of subthreshold scaling is made. The non-native heavy-hex architecture is key to unveiling these results, since native surface code implementations parallelize operations and leave no idling gaps that would lead to this subtle interplay of QEC and DD. The findings presented here constitute a key step in the demonstration of subthreshold surface code scaling in non-native architectures, paving the way for future experiments with near-term QPUs and larger codes. Results Surface code on heavy-hex QPUs. The distance parameter d of an [[n, k, d]] stabilizer code can be characterized as the smallest possible weight of physical Pauli operators appearing in the logical operators XL, ZL38. It is useful to distinguish between error types in this context and, as mentioned above, introduce different distance parameters, e.g., dx and dz for Pauli X and Z errors, respectively. As in the surface code, this distinction can be particularly relevant for CSS codes2,3, as these errors are in practice corrected independently. From a practical standpoint, current Heron QPUs can support a SWAP-based embedding of exactly three odd-distance such surface code sizes: (dx, dz) = (3, 3), (3, 5), and (5, 3) (see Fig. 1), allowing exploration of scaling even though the (5, 5) code is out of reach. Henceforth, we use d ≡ (dx, dz) to denote these codes. The operation of the surface code as a quantum memory involves encoding, syndrome extraction, error decoding, and recovery operations. We next describe the syndrome extraction step, i.e., the measurement of plaquette and vertex stabilizers of the surface code. ref. 25 devised a SWAP-based protocol of this step on the heavy-hex lattice, which was found to be the best performing strategy compared to several other embeddings and promising native codes. Instead of using a naive CNOT-gate compilation of SWAP operations6, which must be applied before and after the syndrome-extraction CNOTs to transfer the state of the non-neighboring data qubits that form a specific stabilizer, the SWAP-based protocol uses a folding-unfolding scheme minimizing circuit depth. The circuits initially follow the middle-out ancilla-free measurement scheme described in ref. 39, and start by applying a first layer of CNOT gates between data qubits to fold weight-4 stabilizers into weight-2 operators. The middle-out protocol would then apply another CNOT layer to further fold these to one-body operators, which can then be measured. Instead, in light of the heavy-hex layout, it is better to use an ancilla qubit to measure the weight-2 parity check directly. Afterward, the weight-2 checks are unfolded back into the original weight-4 stabilizers by a sequence of CNOTs. Denoting A-F as the bulk data qubits (see Fig. 1), the stabilizers ZAZBZCZD and XBXCXEXF are mapped to ZBZC and XEXF during the first step, respectively. Then ZBZC and XEXF are measured using an ancilla qubit. Finally, the original stabilizers are restored (unfolded) by using consecutive CNOTs. This step can be further optimized by executing the first two steps of the following QEC round with the last ones from the previous round (see Fig. 1), minimizing the overall circuit depth25. We further optimize the circuits by removing reset gates, instead tracking the previous measurement outcomes in software40. Removing the resets significantly shortens the duration of the syndrome extraction round, thus also reducing idling errors. As most CNOT gates that implement the SWAP operation cancel out, the result is a compact depth-7 circuit, shown in Fig. 1b. Note that in this scheme, only half of the stabilizers can be measured in parallel, requiring two rounds to perform a full syndrome extraction cycle. We note that ref. 18 ran syndrome extraction circuits by effectively connecting next-nearest-neighbor qubits, which required translating each CNOT gate into a sequence of four consecutive CNOT gates. This contributes to overhead and partly explains why their surface code experiments did not achieve subthreshold behavior. The lack of optimized DD is likely another major reason. A noteworthy aspect of ref. 18 is the interesting idea of recycling the spare qubits for an additional code, allowing for novel fault-tolerant entangling gates. We also note that while the fold-unfold SWAP embedding builds on the theoretical construction of ref. 25, here we develop a hardware-tailored compilation for heavy-hex processors that combines fold-unfold routing, reset elimination, bridge-ancilla usage, and depth minimization under the constraints of the actual device schedule. This enables an experimental multi-cycle surface-code implementation with logical-performance benchmarking and direct DD/no-DD comparisons on current superconducting hardware. Experimental surface code memory with dynamical decoupling. Investigating subthreshold surface code scaling in the heavy-hex lattice amounts to comparing the largest embeddable code with smaller codes restricted to sublattices of the same superconducting chip. In Fig. 1a, we show the layouts used in this work, which include the (5, 3) and (3, 5) surface codes, as well as three patches that can host the (3, 3) code in a sublattice. We note that the (3, 5) and (5, 3) codes grow in different spatial directions, thus involving different sets of transmon qubits and couplers, with varying T1/T2 times and native gate errors. This prevents us from making a fair comparison between them. Instead, we benchmark each anisotropic code against both the average of the three (3, 3) code patches it includes and the top-performing patch. We also note that it is possible to swap the X and Z stabilizers to implement an alternative, conjugate set of codes. As is clear from Fig. 1, this effectively corresponds to reversing the chirality of each code and using the opposite set of qubits, i.e., using the red layout for the (5, 3) code and the green layout for (3, 5). We also performed experiments with these conjugate codes. The results were very similar to those we report below, and are not shown. We perform a quantum memory experiment by initializing each code in four logical states, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert \alpha \right\rangle }$$\end{document}α¯, α ∈ {0, 1, + , − }, which requires preparing the data qubits in either \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\otimes }_{n}\left\vert 0\right\rangle$$\end{document}⊗n0 for α ∈ {0, 1} or \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\otimes }_{n}\left\vert+\right\rangle$$\end{document}⊗n+ for α ∈ { + , − }. In addition, we apply the logical operator XL or ZL to prepare \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\vert 1\right\rangle$$\end{document}1 and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\vert -\right\rangle$$\end{document}−, respectively. We proceed by applying N QEC cycles, each consisting of two consecutive rounds of SWAP-based folding/unfolding circuits for all X- and Z-type stabilizers, yielding a collection of bitstrings that specify which stabilizers have flipped at a particular QEC cycle. This is illustrated in Fig. 1c, d. For the first cycle, which we denote by N = 0, the X-type (for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert 0\right\rangle }$$\end{document}0¯ and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert 1\right\rangle }$$\end{document}1¯) or Z-type (for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert+\right\rangle }$$\end{document}+¯ and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert -\right\rangle }$$\end{document}−¯) stabilizers project the prepared qubits into the code space, and their random results are used to define the initial Pauli frame, as well as to extract and calibrate the SPAM errors [Fig. 1c]. The subsequent cycles [Fig. 1d] project the state into a possibly different stabilizer subspace specified by a bitstring that determines the corresponding ± 1 values of the plaquette and vertex operators, providing an evolving error syndrome that must then be fed to a decoder, as explained in more detail below. This decoder estimates the most likely errors and suggests a recovery correction that can be applied at the software level by a final Pauli-frame update that effectively defines the corrected logical operators. After N such QEC cycles, we projectively measure the data qubits in the corresponding basis, counting logical failures in light of the final Pauli frame for the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert 0\right\rangle }$$\end{document}0¯ and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert 1\right\rangle }$$\end{document}1¯ \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert+\right\rangle }$$\end{document}+¯ and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert -\right\rangle }$$\end{document}−¯) states. This gives access to the resulting X-type (Z-type) logical errors, estimating the error probabilities \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{p}}_{N , 0}$$\end{document}p¯N,0 and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{p}}_{N , 1}$$\end{document}p¯N,1 (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{p}}_{N , +}$$\end{document}p¯N,+ and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{p}}_{N , -}$$\end{document}p¯N,−) through the relative frequencies. We perform 10 QEC cycles, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{{{\min }}}=0$$\end{document}Nmin=0 to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{{{\max }}}=9$$\end{document}Nmax=9, totaling approximately 80μs, which corresponds to 20 embedded folding-unfolding rounds of surface code stabilizer readout. We make heavy use of DD to suppress both coherent and incoherent errors26–29. The main goal is to reduce the burden on the surface code, leaving it to predominantly handle just the incoherent Markovian errors that DD cannot suppress. Although DD during idle periods is well established15,17, here it is integrated in a hardware-aware and gap-aware manner into the full multi-cycle surface code syndrome-extraction circuit. We optimize robustness-screened sequences for the uneven idle windows of both data and ancilla qubits and experimentally compare QEC performance with and without DD, showing that optimized DD is necessary to suppress correlated, slowly varying errors without degrading logical performance. To this end, we optimize performance over a set of DD sequences vetted in the context of IBM QPUs41, and primarily chosen from the robust pulse sequence family known as “universally robust”, or URm32. These sequences are designed to be resilient to pulse imperfections, including over/under-rotation and axis angle errors. The integer m corresponds to the number of pulses, and as long as only the modeled errors occur, DD can reduce the infidelity to ϵm by increasing m, where ϵ is a measure of pulse error. In reality, however, longer sequences of imperfect pulses may accumulate unmodeled errors, which can actually overwhelm the suppression of those that are modeled, leading to overall performance degradation. In addition, in contrast to typical QEC experiments where DD is applied only to idling data qubits during readout and resets, our circuits also include idling bridge and ancillary qubits. Accordingly, we apply DD to all qubits, regardless of their role. However, uniform application of DD across all gaps also degrades performance. Hence, it is important to search for the optimal number of pulses m per sequence, which will typically depend on the specific idle gap in the circuit, as these have different durations and cannot all accommodate a unique DD sequence. We augment the URm family with two other robust sequences for the shorter idle gaps, XY427,42 and RGA8a31, and find that the set {URm, XY4, RGA8a} with m ∈ {6, 8, …, 18} results in the top-performing DD strategy for our surface code quantum memory. We have devised an iterative algorithm for identifying such a strategy that optimizes the filling of idle gaps from this set in order of decreasing time duration. As shown below, this approach results in a substantial improvement in surface code performance relative to not using DD, or to using unoptimized DD variants, such as uniformly filling idle gaps with non-robust CPMG sequences43. The inclusion of DD results in the addition of up to ≈ 484 single-qubit DD pulses per cycle used to fill idle gaps (see ref. 33). Calibrated non-uniform noise model and decoding. Our experiments are decoded using Stim44 and the minimum-weight perfect-matching (MWPM) decoder45, which defines the logical observable for each measurement shot resulting from the corresponding Pauli frame updates. A logical failure is recorded when this observable disagrees with the measured logical observable, as this signals that the quantum memory has been altered. The logical error probability \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{p}}_{N , \alpha }$$\end{document}p¯N,α is then estimated as the fraction of failures over all shots. In order to predict the logical observable, the decoder must be provided with a circuit-level noise model. In our case, this is a multichannel Markovian noise model that includes depolarizing channels for the gates; biased dephasing channels that depend on the T1 and T2 values adapted to the different idling times; and bit-flip channels for the measurements (see ref. 33). The rates of these channels are obtained from the calibration data of each qubit and qubit pair and display a characteristic non-uniformity (rates depend on per-qubit T1/T2, gate fidelities, etc.), as each qubit and coupler acquire different specifications during fabrication. Given the use of optimized, robust DD sequences that suppress pulse angle and axis errors to a certain order, our decoder assumes ideal single-qubit gates. When our decoding includes DD, we report the performance based on the experimentally optimized DD for each data point. Suppression factor metric for assessing subthreshold scaling. To quantify evidence of subthreshold behavior, we begin with the conventional suppression factor between consecutive distances, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Lambda }_{\varepsilon }^{(d)}\equiv \frac{{\varepsilon }_{d}}{{\varepsilon }_{d+2}}$$\end{document}Λε(d)≡εdεd+2, where εd denotes the average logical error probability per QEC cycle at code distance d15,17,19,23,46. Under standard threshold-model assumptions, one expects an approximately exponential decay of the logical error with distance εd ≈ Ce−κd, implying a suppression factor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Lambda }_{\varepsilon }^{(d)}\approx {e}^{2\kappa }$$\end{document}Λε(d)≈e2κ that is approximately constant in d and strictly greater than 1 in the subthreshold regime. Equivalently, a d-independent suppression factor Λε > 1 over a range of distances is consistent with exponential suppression of the per-cycle logical error probability with distance, with rate parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\kappa=\frac{1}{2}{{{\rm{ln}}}}{\Lambda }_{\varepsilon }$$\end{document}κ=12lnΛε. If such behavior persists to larger d, it indicates scalable suppression that, in principle, could enable practically relevant algorithms on large QPUs (megaquop47 or even teraquop48 regimes). In current quantum-memory experiments, reported suppression factors lie in the range \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Lambda }_{\varepsilon }^{(d)} \sim [1.56 , 2.15]$$\end{document}Λε(d)~[1.56,2.15] up to d = 7 17,19. To estimate εd, the typical approach is to fit the error probability \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{p}}_{N , \alpha }^{{{{\boldsymbol{d}}}}}$$\end{document}p¯N,αd as a function of the number of QEC cycles N to a phenomenological single-parameter model 1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{p}}_{N , \alpha }^{d}=\frac{1}{2}-\frac{1}{2}{\left(1-2{\varepsilon }_{\alpha }^{d}\right)}^{N} , $$\end{document}p¯N,αd=12−121−2εαdN,and then proceed by averaging the results over possible sublattices and logical basis states with label α to provide a single per-cycle error rate εd for at least two consecutive code distances17. This single-parameter error model can be analytically derived from a purely combinatorial calculation, assuming that the quantum memory has a constant logical error probability per cycle \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\varepsilon }_{\alpha }^{d}$$\end{document}εαd, and that only odd numbers of logical errors can contribute to a logical failure49. An alternative derivation uses an ad hoc difference equation for the logical error probability update rule after each cycle15. As we discuss in detail below, Eq. (1) is valid only under stationarity (no cycle dependence), unitality (non-increasing purity), and no SPAM errors in the logical error channel. Hence, applying Eq. (1) without ensuring that the model assumptions are satisfied may yield unfaithful results. Entanglement fidelity metric for assessing subthreshold scaling. As mentioned in the Introduction, a first-principles quantum memory metric can be derived from the entanglement fidelity (EF) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F}_{{{{\rm{e}}}}}=\left\langle \phi \right\vert \left({{{\mathcal{I}}}}\otimes \Psi \right)\left(\left\vert \phi \right\rangle \left\langle \phi \right\vert \right)\left\vert \phi \right\rangle$$\end{document}Fe=ϕI⊗Ψϕϕϕ, where Ψ is a quantum channel, i.e., a completely-positive trace-preserving (CPTP) map, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\vert \phi \right\rangle=\frac{1}{{D}^{1/2}}{\sum }_{j=0}^{D-1}\left\vert j\right\rangle \otimes \left\vert j\right\rangle$$\end{document}ϕ=1D1/2∑j=0D−1j⊗j is a maximally entangled state in a D2-dimensional Hilbert space, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\mathcal{I}}}}$$\end{document}I is the identity channel. Fe is directly related to the average channel fidelity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F}_{{{{\rm{ave}}}}}=\int\,d\psi \left\langle \psi \right\vert \Phi (\left\vert \psi \right\rangle \left\langle \psi \right\vert)\left\vert \psi \right\rangle=({F}_{{{{\rm{e}}}}}+1/D)/(1+1/D)$$\end{document}Fave=∫dψψΦ(ψψ)ψ=(Fe+1/D)/(1+1/D), and measures how well the channel preserves an arbitrary state. For a qubit, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F}_{{{{\rm{e}}}}}=\frac{1}{4}{\sum }_{i}| {{{\rm{Tr}}}}({E}_{i}){| }^{2}$$\end{document}Fe=14∑i∣Tr(Ei)∣2, where {Ei} are Ψ’s Kraus operators. For a logical qubit with a Markovian error model described by a CP-divisible map50, such that one simply composes the quantum channels for each cycle to obtain the full noisy evolution, calculating the EF requires estimating the logical error probabilities per cycle of the six cardinal states \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\{\overline{\left\vert 0\right\rangle } , \overline{\left\vert 1\right\rangle } , \overline{\left\vert+\right\rangle } , \overline{\left\vert -\right\rangle } , \overline{\left\vert+{{{\rm{i}}}}\right\rangle } , \overline{\left\vert -{{{\rm{i}}}}\right\rangle }\}$$\end{document}{0¯,1¯,+¯,−¯,+i¯,−i¯}36. However, to date, surface code experiments only give access to logical error probabilities in the X and Z basis \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\{{\overline{p}}_{N , 0}^{{{\boldsymbol{d}}}} , {\overline{p}}_{N , 1}^{{{\boldsymbol{d}}}} , {\overline{p}}_{N , + }^{{{\boldsymbol{d}}}} , {\overline{p}}_{N , - }^{{{\boldsymbol{d}}}}\}$$\end{document}{p¯N,0d,p¯N,1d,p¯N,+d,p¯N,−d}, corresponding to logical X and Z errors, respectively, since Y basis preparation and readout require significantly more complicated methods35. Nevertheless, under reasonable assumptions about the logical error channel, we can derive compact expressions for the EF that go beyond a simple averaging of the basis-dependent logical errors and approximate the contributions from the Y-basis states. Our model describes the logical quantum channel as consisting of independent logical Pauli X and Z errors, logical amplitude damping, and logical SPAM errors, and assumes a CP-divisible (cycle-composable) logical channel across cycles. Under this model, we derive expressions for the EF \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F}_{{{{\rm{e}}}}}^{{{{\boldsymbol{d}}}}}(N)$$\end{document}Fed(N), as well as rigorous upper and lower bounds so that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F}_{{{{\rm{e}}}} , < }^{{{{\boldsymbol{d}}}}}(N)\le {F}_{{{{\rm{e}}}}}^{{{{\boldsymbol{d}}}}}(N)\le {F}_{{{{\rm{e}}}} , \ > \ }^{{{{\boldsymbol{d}}}}}(N)$$\end{document}Fe,d(N), which solely use the decoder-provided (measured) probabilities (see ref. 33): 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F}_{{{{\rm{e}}}}}^{{{{\boldsymbol{d}}}}}(N)={F}_{{{{\rm{e}}}} , < }^{{{{\boldsymbol{d}}}}}(N)+\left(\frac{1}{{\Gamma }^{(N)}}-1\right)r(N) , $$\end{document}Fed(N)=Fe, \ }^{{{{\boldsymbol{d}}}}}(N)={F}_{{{{\rm{e}}}} , < }^{{{{\boldsymbol{d}}}}}(N)+\left(\frac{{\Sigma }_{z}^{{{{\boldsymbol{d}}}}}(0)}{{\Sigma }_{z}^{{{{\boldsymbol{d}}}}}(N)}-1\right)r(N) , $$\end{document}Fe,>d(N)=Fe, {\tilde{{F}_{{{{\rm{e}}}}}}}^{(d^{\prime}_{x} , d^{\prime}_{z})}(N)$$\end{document}Fe~(dx,dz)(N)>Fe~(dx′,dz′)(N) for all N, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${d}_{x} > d^{\prime} _{x}$$\end{document}dx>dx′ or \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${d}_{z} > d^{\prime}_{z}$$\end{document}dz>dz′, or if the ratio of infidelities 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Lambda }_{F}^{({{{\boldsymbol{d}}}} , {{{{\boldsymbol{d}}}}}^{{\prime} })}(N)\equiv \frac{1-{\tilde{F}}_{{{{\rm{e}}}}}^{{{{{\boldsymbol{d}}}}}^{{\prime} }}(N)}{1-{\tilde{F}}_{{{{\rm{e}}}}}^{{{{\boldsymbol{d}}}}}(N)}$$\end{document}ΛF(d,d′)(N)≡1−F~ed′(N)1−F~ed(N)satisfies \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Lambda }_{F}^{({{{\boldsymbol{d}}}} , {{{\boldsymbol{d}}}}^\prime)}(N) > 1$$\end{document}ΛF(d,d′)(N)>1 for all N under the same distance ordering, then this constitutes evidence of subthreshold scaling. Maximization over the inclusion or exclusion of DD ensures that each code is compared in its optimally configured form, an important point that we highlighted in the Introduction and discuss in more detail below. In applying Eq. (4) in the context of individual sublattices, we define the top-performing sublattice code as the code that has the highest fidelity averaged over all cycles. This is consistent with keeping the logical qubit fidelity as high as possible for as long as possible. We advocate for the use of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Lambda }_{F}^{({{{\boldsymbol{d}}}} , {{{\boldsymbol{d}}}}^\prime)}(N)$$\end{document}ΛF(d,d′)(N) in lieu of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Lambda }_{\varepsilon }^{(d)}$$\end{document}Λε(d), as it is a fitting-model-free metric that, moreover, does not require the assumption of stationarity and can be computed directly from the logical error probabilities output by the decoder. In fact, it is by carefully considering the EF that we can identify the implicit assumptions underlying the suppression factor, and the required generalizations of fitting error models and logical basis averaging. Reduction to fitting models. As shown in ref. 33, when we assume (1) stationarity (the noise channel is independent of the cycle number N), (2) unitality (the noise consists purely of Pauli errors), and (3) no logical SPAM errors, the single-parameter error model Eq. (1) is recovered from first principles. The EF in this case coincides exactly with the lower bound for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Sigma }_{x}^{{{{\boldsymbol{d}}}}}(0)={\Sigma }_{z}^{{{{\boldsymbol{d}}}}}(0)=1$$\end{document}Σxd(0)=Σzd(0)=1 and, after expanding the EF to leading order in the logical error probabilities, gives \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F}_{{{{\rm{e}}}}}^{{{{\boldsymbol{d}}}}}(N)=\frac{1}{4}\left([1+{\Sigma }_{x}^{{{{\boldsymbol{d}}}}}(N)][1+{\Sigma }_{z}^{{{{\boldsymbol{d}}}}}(N)]\right)\approx 1-{p}_{N}^{{{{\boldsymbol{d}}}}}$$\end{document}Fed(N)=14[1+Σxd(N)][1+Σzd(N)]≈1−pNd, with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{p}}_{N}^{{{{\boldsymbol{d}}}}}=\frac{1}{4}{\sum }_{\alpha }{\overline{p}}_{N , \alpha }^{{{{\boldsymbol{d}}}}}$$\end{document}p¯Nd=14∑αp¯N,αd as the logical error probability after N cycles averaged over the four basis states. Note that each of these error probabilities is described by the single-parameter error model Eq. (1), and that only for N = 1 one recovers an expression with the per-cycle error rates averaged over basis states \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F}_{{{{\rm{e}}}}}^{{{{\boldsymbol{d}}}}}(1)=1-{\epsilon }_{{{{\boldsymbol{d}}}}}$$\end{document}Fed(1)=1−ϵd, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\epsilon }_{{{{\boldsymbol{d}}}}}=\frac{1}{4}{\sum }_{\alpha }{\epsilon }_{\alpha }^{{{{\boldsymbol{d}}}}}$$\end{document}ϵd=14∑αϵαd. For generic N, it is not the averaged ϵd which determines the EF and the average channel infidelity, but a linear combination of the different exponential decays per basis state, in addition to non-linear contributions due to the Y-basis states. When the specific assumptions above do not hold, the complete expression Eq. (2) gives a more informative metric. The non-stationary, non-unital, SPAM-inclusive model used to derive Eq. (2) is actually crucial in our case, as our QEC experiments show evidence of cycle-dependence as well as logical amplitude damping and SPAM errors. The same first-principles model can also be used to derive a generalization of Eq. (1) with three free but stationary (cycle-independent) parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\{{a}_{\alpha } , {b}_{\alpha } , {\varepsilon }_{\alpha }^{{{{\boldsymbol{d}}}}}\}$$\end{document}{aα,bα,εαd} that can be used to fit the logical error probability 5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{p}}_{N , \alpha }^{{{{\boldsymbol{d}}}}}={a}_{\alpha }+{b}_{\alpha }{\left(1-{\varepsilon }_{\alpha }^{{{{\boldsymbol{d}}}}}/{a}_{\alpha }\right)}^{N} , $$\end{document}p¯N,αd=aα+bα1−εαd/aαN,where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\varepsilon }_{\alpha }^{{{{\boldsymbol{d}}}}}$$\end{document}εαd is the per-cycle error rate (a probability) and aα, bα account for non-unital and SPAM errors. Interestingly, the same result can also be derived from a physically motivated difference equation for the logical error probabilities, relating the free parameters to specific logical error sources (see ref. 33). The aforementioned reduction to the one-parameter model Eq. (1) by imposing additional assumptions is particularly clear when Eq. (5) is used as the starting point: neglecting SPAM errors sets bα = − aα, and restricting to unital Pauli errors further fixes aα = 1/2. For the IBM QPUs, we report different values for aα and bα below as required for a best-fit to Eq. (5), attesting that the logical error model is not a simple Pauli channel. Logical error probabilities. We now present our results for the SWAP-based embedding of the surface code using ibm_aachen, the top-performing QPU in our study, and its integration with robust DD. Recall that the raw decoder outputs are the logical error probabilities after N QEC cycles, i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{p}}_{N , \alpha }^{{{{\boldsymbol{d}}}}}$$\end{document}p¯N,αd from experiments initialized in the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert \alpha \right\rangle }$$\end{document}α¯ state with α ∈ {0, 1, + , − }, which we analyze in light of the metrics discussed above. We set N = 0 to denote the SPAM-calibration cycle. Since the (3, 5) code increases protection against phase-flip (Z) errors and the (5, 3) code increases protection against bit-flip (X) errors, we analyze (i) basis-resolved \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{p}}_{N , \alpha }^{{{{\boldsymbol{d}}}}}$$\end{document}p¯N,αd, (ii) the basis-specific suppression factor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Lambda }_{\varepsilon , \alpha }^{{{{\boldsymbol{d}}}}}$$\end{document}Λε,αd extracted from fits, and (iii) a basis-independent entanglement-fidelity (EF) metric \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F}_{{{{\rm{e}}}}}^{{{{\boldsymbol{d}}}}}(N)$$\end{document}Fed(N) and its ratio ΛF defined in Eq. (4). Among prior quantum memory experiments, ref. 51 prepared all four states; Refs. 15,17,22 used only α ∈ {0, + }; ref. 19 also prepared α ∈ {1, − } for neural-network decoder training but did not report memory results for these states. Figure 2 shows our experimental logical error probabilities \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{p}}_{N , \alpha }^{{{{\boldsymbol{d}}}}}$$\end{document}p¯N,αd versus the number of QEC cycles N for the representative states α ∈ {0, + }. These probabilities reflect control, measurement, and environment-induced errors, as well as decoder accuracy. For each data point, we report the minimum \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{p}}_{N , \alpha }^{{{{\boldsymbol{d}}}}}$$\end{document}p¯N,αd found over a search of decoder noise-model parameters and DD configurations. Because the (3, 3) code can be embedded in three distinct sublattices, we report \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{p}}_{N , \alpha , s}^{(3 , 3)}$$\end{document}p¯N,α,s(3,3) for each sublattice s, along with their average \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{p}}_{N , \alpha }^{(3 , 3)}=\frac{1}{3}{\sum }_{s=1}^{3}{\overline{p}}_{N , \alpha , s}^{(3 , 3)}$$\end{document}p¯N,α(3,3)=13∑s=13p¯N,α,s(3,3). Figure 2a shows that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{p}}_{N , +}^{(3 , 5)} < {\overline{p}}_{N , +}^{(3 , 3)}$$\end{document}p¯N,+(3,5) 1 indicate basis-specific subthreshold improvement. The bottom row reports \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Lambda }_{\varepsilon }^{{{{\boldsymbol{d}}}}}=\frac{{\varepsilon }^{(3,3)}}{{\varepsilon }^{{{{\boldsymbol{d}}}}}}$$\end{document}Λεd=ε(3,3)εd where (3, 3) is the average over the four initial states of the three sublattice codes, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\varepsilon }^{{{{\boldsymbol{d}}}}}=\frac{1}{4}{\sum }_{\alpha \in \{+, - , 0, 1\}}{\varepsilon }_{\alpha }^{{{{\boldsymbol{d}}}}}$$\end{document}εd=14∑α∈{+,−,0,1}εαd. All uncertainties (omitted) are < 2 × 10−3. Table 1Basis-specific suppression factors \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Lambda }_{\varepsilon,\alpha }^{{{{\boldsymbol{d}}}}}$$\end{document}Λε,αd (scaled code vsSuppressiond = (5, 3)d = (3, 5)factor metricα = 0α = 1α = + α = − \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Lambda }_{\varepsilon,\alpha }^{{{{\boldsymbol{d}}}}}$$\end{document}Λε,αd1.461.461.231.24\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathop{\min }_{s}}{\Lambda }_{\varepsilon,\alpha,s}^{{{{\boldsymbol{d}}}}}$$\end{document}minsΛε,α,sd1.101.071.041.01\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Lambda }_{\varepsilon }^{{{{\boldsymbol{d}}}}}$$\end{document}Λεd0.670.77sublattice-average (3, 3)) and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\min }_{s}}{\Lambda }_{\varepsilon,\alpha,s}^{{{{\boldsymbol{d}}}}}$$\end{document}minsΛε,α,sd (scaled code vs. best (3, 3) sublattice). Values > 1 indicate basis-specific subthreshold improvement. The bottom row reports \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Lambda }_{\varepsilon }^{{{{\boldsymbol{d}}}}}=\frac{{\varepsilon }^{(3,3)}}{{\varepsilon }^{{{{\boldsymbol{d}}}}}}$$\end{document}Λεd=ε(3,3)εd where (3, 3) is the average over the four initial states of the three sublattice codes, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\varepsilon }^{{{{\boldsymbol{d}}}}}=\frac{1}{4}{\sum }_{\alpha \in \{+, - , 0, 1\}}{\varepsilon }_{\alpha }^{{{{\boldsymbol{d}}}}}$$\end{document}εd=14∑α∈{+,−,0,1}εαd. All uncertainties (omitted) are < 2 × 10−3. Basis-specific suppression factors \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Lambda }_{\varepsilon,\alpha }^{{{{\boldsymbol{d}}}}}$$\end{document}Λε,αd (scaled code vs sublattice-average (3, 3)) and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\min }_{s}}{\Lambda }_{\varepsilon,\alpha,s}^{{{{\boldsymbol{d}}}}}$$\end{document}minsΛε,α,sd (scaled code vs. best (3, 3) sublattice). Values > 1 indicate basis-specific subthreshold improvement. The bottom row reports \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Lambda }_{\varepsilon }^{{{{\boldsymbol{d}}}}}=\frac{{\varepsilon }^{(3,3)}}{{\varepsilon }^{{{{\boldsymbol{d}}}}}}$$\end{document}Λεd=ε(3,3)εd where (3, 3) is the average over the four initial states of the three sublattice codes, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\varepsilon }^{{{{\boldsymbol{d}}}}}=\frac{1}{4}{\sum }_{\alpha \in \{+, - , 0, 1\}}{\varepsilon }_{\alpha }^{{{{\boldsymbol{d}}}}}$$\end{document}εd=14∑α∈{+,−,0,1}εαd. All uncertainties (omitted) are < 2 × 10−3. Although averaging over sublattices is motivated by the fact that all the corresponding qubits and couplers will also be used in the larger codes, when solving a computational problem of a given size, it is more natural to aim for the best possible performance, i.e., to optimize over the available sublattices. We find that the improvement holds even from this perspective. Namely, with respect to the best sublattice we obtain \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\min }_{s}}{\Lambda }_{\varepsilon , \alpha , s}^{{{{\boldsymbol{d}}}}}\in [1.01 , 1.10]$$\end{document}minsΛε,α,sd∈[1.01,1.10]. This shows, for the first time, that current IBM QPUs support subthreshold scaling, albeit only with respect to anisotropic growth of the code distance; we clarify below why a global subthreshold advantage is not yet achieved. With these caveats noted, to the best of our knowledge—apart from the results of ref. 15 for roughly the same amount of QPU time (~ 80μs)—this is currently the only other demonstration of subthreshold scaling in superconducting-qubit QPUs. We also report in Table 1 the suppression factor obtained after averaging over all four input states, namely \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Lambda }_{\varepsilon }^{{{{\boldsymbol{d}}}}}=\frac{{\varepsilon }^{(3 , 3)}}{{\varepsilon }^{{{{\boldsymbol{d}}}}}}$$\end{document}Λεd=ε(3,3)εd where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\varepsilon }^{(3 , 3)}=\frac{1}{4}{\sum }_{\alpha \in \{ +, - , 0 , 1\}}{\varepsilon }_{\alpha }^{(3 , 3)}$$\end{document}ε(3,3)=14∑α∈{+,−,0,1}εα(3,3) and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\varepsilon }^{{\boldsymbol{d}}} =\frac{1}{4}{\sum }_{\alpha \in \{+, -, 0 , 1\}}{\varepsilon }_{\alpha }^{{\boldsymbol{d}}}$$\end{document}εd=14∑α∈{+,−,0,1}εαd, with d ∈ {(3, 5), (5, 3)}. This suppression factor lies well below 1 for both scaled codes, suggesting that the anisotropic scaling does not suffice to provide increased protection for all basis states. In other words, the improvement due to QEC in the logical basis parallel to the growth direction does not overcome the degradation of the orthogonal basis caused by the increased circuit complexity without any further protection. However, this basis-state averaging is ad hoc (see ref. 33) and the entire metric is susceptible to model-selection sensitivity. To illustrate the latter, we recomputed \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Lambda }_{\varepsilon }^{{{{\boldsymbol{d}}}}}$$\end{document}Λεd under the one- and two-parameter fits. The resulting spread, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left(\max {\Lambda }_{\varepsilon }^{{{{\boldsymbol{d}}}}}-\min {\Lambda }_{\varepsilon }^{{{{\boldsymbol{d}}}}}\right)/{{{\rm{mean}}}}\,{\Lambda }_{\varepsilon }^{{{{\boldsymbol{d}}}}}$$\end{document}maxΛεd−minΛεd/meanΛεd, reaches 14.93% for (3, 5) with a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert+\right\rangle }$$\end{document}+¯ memory and 17.22% for (5, 3) with a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert 0\right\rangle }$$\end{document}0¯ memory. This underscores the need for AIC-guided model choice rather than arbitrary parameter-number selection. Entanglement fidelity QEC metric. So far, we have mostly considered the effect of code scaling on the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\{\overline{\left\vert 0\right\rangle } , \overline{\left\vert 1\right\rangle } , \overline{\left\vert+\right\rangle } , \overline{\left\vert -\right\rangle }\}$$\end{document}{0¯,1¯,+¯,−¯} states separately. For the scaling of the suppression factor, this is motivated by the fact that different fitting parameters are required for each of the states [recall Eq. (5)]. However, a genuine quantum memory should exhibit subthreshold scaling independent of the input state, and to assess this, we must use an unambiguous metric that inherently combines the results for the different basis states. A common approach based on the logical error rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\varepsilon }_{\alpha }^{{{{\boldsymbol{d}}}}}$$\end{document}εαd or the suppression factor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Lambda }_{\varepsilon , \alpha }^{{{{\boldsymbol{d}}}}}$$\end{document}Λε,αd, is to average over the basis states17. As explained in ref. 33, this approach can only be rigorously connected to the channel infidelity of the quantum memory after a single N = 1 QEC cycle, and in the limit of small logical Pauli errors. For larger cycle numbers, a more complicated expression is needed that would, in addition, require the measurement of all six cardinal states. The natural alternative to the basis-dependent approach is the EF, which, as we showed above, combines all the data provided by the decoder into a single, easily computable and rigorous performance metric. Using the EF [Eq. (2)] removes the ambiguity associated with model fitting and, moreover, accounts for SPAM and non-unital errors (as motivated by our three-parameter model), as well as cycle-dependence. We plot the resulting entanglement infidelities in Fig. 3, and their ratios, i.e., the entanglement fidelity metric of Eq. (4), in Fig. 4.Fig. 3Entanglement fidelity analysis.Entanglement infidelity 1 − Fe(N) for (a) (3, 5) and (b) (5, 3). As in Fig. 2, blue open markers are the two sets of three (3, 3) sublattices relevant to each scaled code. Filled blue: arithmetic averages over those sublattices. We set 1 − Fe(0) = 0 by removing logical SPAM; Fe(N) is computed from X/Z data using Eq. (2).Fig. 4Surface code scaling anlysis.Optimized EF metric ΛF [Eq. (4)] relative to the (3, 3) sublattice average (black with error bars) and relative to the best (3, 3) sublattice (purple, shaded). Also shown is the unoptimized EF metric \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{1-{F}_{{{{\rm{e}}}}}^{{{{\boldsymbol{d}}}}^\prime }(N)}{1-{F}_{{{{\rm{e}}}}}^{{{{\boldsymbol{d}}}}}(N)}$$\end{document}1−Fed′(N)1−Fed(N) without DD (green, shaded). All results are for the (5, 3) code and its sublattices implemented on (a) ibm_aachen and (b) ibm_marrakesh. Values > 1 would indicate basis-independent subthreshold scaling. Consistent with Fig. 3, this is not observed for ibm_aachen once DD is considered. For ibm_marrakesh, we observe the onset of subthreshold scaling for N ≥ 5 in terms of both the sublattice-average and the best sublattice, but with low statistical confidence. The no DD (5, 3) case exceeds 1 for both ibm_aachen and ibm_marrakesh but is spurious, since the correct reference compares DD-optimized \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\max }_{\{\,{\mbox{DD}},{\mbox{noDD}}\,\}}}{F}_{{{{\rm{e}}}}}$$\end{document}max{DD,noDD}Fe for each code, as per Eq. (4). To select the best sublattice, we compute \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{1}{9}\mathop{\sum }_{N=1}^{9}{F}_{{{{\rm{e}}}},s}^{(3,3)}(N)$$\end{document}19∑N=19Fe,s(3,3)(N) and maximize over s ∈ {1, 2, 3}. Entanglement fidelity analysis. Entanglement infidelity 1 − Fe(N) for (a) (3, 5) and (b) (5, 3). As in Fig. 2, blue open markers are the two sets of three (3, 3) sublattices relevant to each scaled code. Filled blue: arithmetic averages over those sublattices. We set 1 − Fe(0) = 0 by removing logical SPAM; Fe(N) is computed from X/Z data using Eq. (2). Surface code scaling anlysis. Optimized EF metric ΛF [Eq. (4)] relative to the (3, 3) sublattice average (black with error bars) and relative to the best (3, 3) sublattice (purple, shaded). Also shown is the unoptimized EF metric \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{1-{F}_{{{{\rm{e}}}}}^{{{{\boldsymbol{d}}}}^\prime }(N)}{1-{F}_{{{{\rm{e}}}}}^{{{{\boldsymbol{d}}}}}(N)}$$\end{document}1−Fed′(N)1−Fed(N) without DD (green, shaded). All results are for the (5, 3) code and its sublattices implemented on (a) ibm_aachen and (b) ibm_marrakesh. Values > 1 would indicate basis-independent subthreshold scaling. Consistent with Fig. 3, this is not observed for ibm_aachen once DD is considered. For ibm_marrakesh, we observe the onset of subthreshold scaling for N ≥ 5 in terms of both the sublattice-average and the best sublattice, but with low statistical confidence. The no DD (5, 3) case exceeds 1 for both ibm_aachen and ibm_marrakesh but is spurious, since the correct reference compares DD-optimized \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\max }_{\{\,{\mbox{DD}},{\mbox{noDD}}\,\}}}{F}_{{{{\rm{e}}}}}$$\end{document}max{DD,noDD}Fe for each code, as per Eq. (4). To select the best sublattice, we compute \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{1}{9}\mathop{\sum }_{N=1}^{9}{F}_{{{{\rm{e}}}},s}^{(3,3)}(N)$$\end{document}19∑N=19Fe,s(3,3)(N) and maximize over s ∈ {1, 2, 3}. Before interpreting these results, we note that the basis states \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\{\overline{\left\vert 0\right\rangle } , \overline{\left\vert 1\right\rangle }\}$$\end{document}{0¯,1¯} are less protected for the (3, 5) code than \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\{\overline{\left\vert+\right\rangle } , \overline{\left\vert -\right\rangle }\}$$\end{document}{+¯,−¯} (since its smaller dx implies less protection against bit-flip than phase-flip errors), and the same argument holds for the (5, 3) code with interchanged roles for X and Z. Thus, we expect these less protected states to contribute higher logical error probabilities to the EF compared to the (3, 3) code, as we increase the circuit complexity by growing the codes anisotropically. It is unclear a priori whether the advantage gained by growing the code distance for the more protected basis states will overcome the degradation of the EF due to the less protected states. As shown in Figs. 3, 4, this is not the case: the infidelity of the scaled codes is generally higher than that of the smaller codes, such that there is no overall subthreshold scaling. This holds with DD and also without DD, except for the (5, 3) code, which we discuss below. We stress that these results do not depend on a fitting model and, in contrast to an average per-cycle error ϵd and the corresponding suppression factor, are informative at the cycle level well beyond N = 1. However, the conditions for subthreshold scaling according to the full EF metric are actually not far from current performance. In fact, using our circuit-level noise model and numerical simulations for the (5, 5) code, we find (see ref. 33) that a 30% reduction of the current experimental noise rates in a slightly larger QPU that has the six missing qubits needed to host this larger code would already achieve genuine subthreshold scaling, i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F}_{{{{\rm{e}}}}}^{(5 , 5)}(N) > {F}_{{{{\rm{e}}}}}^{(3 , 3)}(N)$$\end{document}Fe(5,5)(N)>Fe(3,3)(N). In ref. 33 we discuss the application of our methodology to the Willow-processor results of ref. 17. We reanalyze the public Willow data with the EF metric, and confirm subthreshold scaling d: 3 → 5 → 7, with a deeper per-cycle advantage for d: 5 → 7 than d: 3 → 5. In contrast, basis-averaged, single-parameter suppression factors vary at the 10–15% level with the chosen fit and presuppose stationarity that is not supported by the data. We therefore advocate EF as a primary, fit-free metric for an unambiguous assessment of subthreshold scaling across different QPUs with different error mechanisms. Role of DD and spurious subthreshold scaling. As mentioned above, our results require careful integration of optimized DD into embedded surface code circuits. Figure 3 also contains a comparison between circuits that do or do not incorporate DD, and it is clear at a glance that, in general, DD significantly reduces the entanglement infidelity. Thus, DD plays a critical role in improving the performance of the surface code SWAP-based embedding in our experiments. Figure 4 shows that for sufficiently large N, the fidelity metric exceeds 1 in the case of the (5, 3) code without DD (green, shaded). This might seem to suggest that subthreshold scaling has been achieved. However, it is essential to first examine the actual entanglement infidelity values and to ensure that they are minimized. In doing so using Fig. 3, we observe that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1-{F}_{{{{\rm{e}}}}}^{(3 , 3)}$$\end{document}1−Fe(3,3) is significantly lower with DD than without it. This means that the correct reference is the case with DD, and a conclusion of subthreshold scaling based on a comparison with the noDD reference is spurious. In other words, as already explained in the context of Eq. (4), each code must be optimized separately with respect to the inclusion of DD. When this is done (black, error bars), we observe that for ibm_aachen, the subthreshold scaling result disappears. For ibm_marrakesh, we do observe subthreshold scaling for N ≥ 6, but not until the last cycle (N = 9) does this result hold with 95% confidence. This represents the onset of subthreshold scaling, i.e., the scaled (5, 3) code gradually becomes more effective than the (3, 3) code as more errors accumulate. Results for the (3, 5) code are well above threshold and are shown in ref. 33, along with a complete set of results for ibm_marrakesh. To further quantify the role played by DD, Fig. 5 shows the improvement factor we obtain by comparing QEC with and without DD, namely (a) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Lambda }_{{{{\rm{DD}}}}}^{({{{\boldsymbol{d}}}})}(N)\equiv {\overline{p}}_{N , \alpha }^{{{{\boldsymbol{d}}}} , {{{\rm{noDD}}}}}/{\overline{p}}_{N , \alpha }^{{{{\boldsymbol{d}}}}}$$\end{document}ΛDD(d)(N)≡p¯N,αd,noDD/p¯N,αd where the noDD superscript indicates that the corresponding circuits were run without any DD pulses, as well as (b) the SPAM-free entanglement infidelity ratios \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Lambda }_{F , {{{\rm{DD}}}}}^{({{{\boldsymbol{d}}}})}(N)\equiv \frac{1-{F}_{{{{\rm{e}}}}}^{{{{\boldsymbol{d}}}} , {{{\rm{noDD}}}}}(N)}{1-{F}_{{{{\rm{e}}}}}^{{{{\boldsymbol{d}}}}}(N)}$$\end{document}ΛF,DD(d)(N)≡1−Fed,noDD(N)1−Fed(N). As seen in Fig. 5a, while the effect on the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert 0\right\rangle }$$\end{document}0¯ state is small (symbols with error bars), the logical error probability of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert+\right\rangle }$$\end{document}+¯ state (symbols with shading) is significantly higher without DD for all codes. This is evidence for a strong bias towards logical Z-type errors, including coherent errors due to residual ZZ crosstalk from tunable couplers, and non-Markovian correlated dephasing. DD strongly suppresses these effects, with a particularly pronounced effect for the (3, 3) code (purple) relative to its (5, 3) extension (brown). The same observation holds after SPAM removal and when accounting for all four basis states, as seen in terms of the infidelity ratios in Fig. 5(b). Figure 5 also shows that the (3, 5) code benefits more from DD than the (5, 3) code. This is consistent with DD’s stronger suppression of Z-type errors. As for the overall decline in the DD improvement factor with N, we attribute this to the aforementioned accumulation of coherent errors (see ref. 33 for more details).Fig. 5Role of dynamical decoupling.The DD improvement factor in terms of (a) ratios of logical error probabilities and (b) entanglement infidelities. In more detail, in (a) we plot \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{p}}_{N,\alpha }^{{{{\boldsymbol{d}}}},{{{\rm{noDD}}}}}/{\overline{p}}_{N,\alpha }^{{{{\boldsymbol{d}}}}}$$\end{document}p¯N,αd,noDD/p¯N,αd, for the scaled codes and their respective (3, 3) code sublattice averages. A value above 1 indicates that DD helps. The shaded confidence intervals are for the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert+\right\rangle }$$\end{document}+¯ state and the error bars for the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert 0\right\rangle }$$\end{document}0¯ state. DD results in a significant improvement in the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert+\right\rangle }$$\end{document}+¯ case, and has essentially no effect in the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert 0\right\rangle }$$\end{document}0¯ case. Results for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert -\right\rangle }$$\end{document}−¯ (significant improvement) and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert 1\right\rangle }$$\end{document}1¯ (no improvement) are nearly identical and are not shown. In (b) we plot \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{1-{F}_{{{{\rm{e}}}}}^{{{{\boldsymbol{d}}}},\,{\mbox{noDD}}\,}(N)}{1-{F}_{{{{\rm{e}}}}}^{{{{\boldsymbol{d}}}}}(N)}$$\end{document}1−Fed,noDD(N)1−Fed(N), which is the SPAM-free version of (a), and also simultaneously accounts for all four basis states. The (3, 5) code (green) is improved more by DD than the (5, 3) code (brown), while the opposite is true for their respective (3, 3) codes. Role of dynamical decoupling. The DD improvement factor in terms of (a) ratios of logical error probabilities and (b) entanglement infidelities. In more detail, in (a) we plot \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{p}}_{N,\alpha }^{{{{\boldsymbol{d}}}},{{{\rm{noDD}}}}}/{\overline{p}}_{N,\alpha }^{{{{\boldsymbol{d}}}}}$$\end{document}p¯N,αd,noDD/p¯N,αd, for the scaled codes and their respective (3, 3) code sublattice averages. A value above 1 indicates that DD helps. The shaded confidence intervals are for the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert+\right\rangle }$$\end{document}+¯ state and the error bars for the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert 0\right\rangle }$$\end{document}0¯ state. DD results in a significant improvement in the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert+\right\rangle }$$\end{document}+¯ case, and has essentially no effect in the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert 0\right\rangle }$$\end{document}0¯ case. Results for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert -\right\rangle }$$\end{document}−¯ (significant improvement) and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\left\vert 1\right\rangle }$$\end{document}1¯ (no improvement) are nearly identical and are not shown. In (b) we plot \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{1-{F}_{{{{\rm{e}}}}}^{{{{\boldsymbol{d}}}},\,{\mbox{noDD}}\,}(N)}{1-{F}_{{{{\rm{e}}}}}^{{{{\boldsymbol{d}}}}}(N)}$$\end{document}1−Fed,noDD(N)1−Fed(N), which is the SPAM-free version of (a), and also simultaneously accounts for all four basis states. The (3, 5) code (green) is improved more by DD than the (5, 3) code (brown), while the opposite is true for their respective (3, 3) codes. Discussion We implemented a surface-code quantum memory on IBM heavy-hex QPUs using a lightweight SWAP-based embedding integrated with robust dynamical decoupling (DD). With this hardware-aware co-design, we observe directional suppression of logical errors under anisotropic distance scaling: increasing dx (dz) improves protection against X-type (Z-type) logical errors, despite increases in circuit width and depth. To benchmark code scaling without model-fit ambiguities, we introduced an EF metric that (i) combines all four accessible basis states; (ii) accounts for SPAM and non-unital effects; and (iii) remains valid under cycle-dependent noise. Using EF, we did not find global, state-independent subthreshold scaling for the full logical channel on current IBM QPUs, with the exception of the onset of such scaling at late QEC cycles for ibm_marrakesh. Our results clarify the limits of inferences drawn from widely used single-parameter, basis-averaged suppression factors, whose assumptions (cycle-independence, unitality, no SPAM) are not fully met in our experiments. DD is essential in this setting. It suppresses dominant Z-biased mechanisms, including coherent ZZ crosstalk and correlated dephasing, thus pushing the residual noise closer to what the surface code can effectively correct. We showed that threshold assessment must optimize each code instance, before comparison, over using or not using DD: using an unoptimized reference can yield spurious indications of subthreshold scaling. In practice, our recommended metric is the entanglement infidelity ratio computed from the best configuration per code. We stress that identifying the best decoder is beyond the scope of the present work: while more accurate decoders than MWPM could plausibly improve the observed scaling under current hardware conditions, a systematic decoder study and its interplay with the improvements from DD is a substantial task that will be addressed in separate work. We also note that an interesting future direction would be to integrate the present experiments with logical-level DD54. However, such integration with the present fold-unfold implementation is nontrivial because the physical representatives of stabilizers and normalizers become cycle-dependent under the SWAP schedule. Incorporating logical-level DD would require tracking these operators through the routing circuit while maintaining compatibility with stabilizer measurements, ancilla resets, readout windows, and the robust DD optimization used to suppress the dominant coherent and low-frequency noise. Circuit-level simulations calibrated to experiment indicate that a modest, ~ 30% reduction in current noise rates, together with access to a slightly larger heavy-hex device that can host a (5, 5) code, would achieve genuine subthreshold scaling under the EF metric. Our results establish a practical path for the demonstration of surface code scaling: (i) minimize depth via connectivity-aware embedding; (ii) integrate robust DD to suppress non-Markovian and coherent error components; and (iii) evaluate scaling using EF-based, SPAM-aware metrics. This combination delivers concrete, device-level targets for realizing below-threshold operation on non-native surface-code architectures, and provides a model-free benchmark for future progress toward fault-tolerant quantum computing. Methods For a detailed description of the methods used, see the Supplementary Information in ref. 33. Supplementary information Supplementary Information Transparent Peer Review file Supplementary Information Transparent Peer Review file