{"id":"b5f0a674-7cfa-456f-88d7-1b1a23ed5107","arxiv_id":"2412.15751","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"On heavy-hexagon hardware with flag qubits, the ZXXZ orientation of the XZZX code with down-triangle qubit initialization gives the lowest logical error rate for magic state injection under biased noise.","lead":"The authors simulate how to prepare 'magic states', special helper states needed for fault-tolerant quantum computing, on IBM-style heavy-hexagon hardware that uses extra flag qubits, and compare with the standard square-lattice design. They find that the extra flag qubits add noise that changes which error-correcting code and initialization pattern works best, and they propose a specific combination for this hardware.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The ZXXZ-vs-XZZX recommendation hinges on the flag-qubit CNOT error-propagation and cancellation model in Sec. III.A and Figs. 5-6; if that model is wrong, the extra-bias mechanism and the recommended orientation could reverse.","rationale":"I read the claim as a simulation-based design recommendation. The numerical trends are plausible, and the XZZX code's known bias resilience makes the qualitative direction sensible. However, the distinguishing feature of the heavy-hex implementation is the flag-qubit network, and the entire orientation preference is explained by how flag errors propagate through that network. The paper's schematic treatment leaves this as an unverified assumption; the reader identified the same point, and I agree. Secondary limitations, including the absence of error bars and shipped code, do not by themselves overturn the simulation, but they make an independent re-implementation the right check. A conditional accept remains appropriate: the recommendation should not be treated as a hardware design rule until the flag-propagation model is validated. Therefore I recommend leaving the reader's verdict unchanged.","tokens_in":33159,"tokens_out":10531,"duration_ms":98005,"concrete_test":"Use Stim to implement the exact heavy-hex stabilizer-measurement circuits for the surface, XZZX, and ZXXZ codes as implied by Figures 2, 3, and 5, reconstructing the circuits from the authors if necessary. Insert each single-qubit Pauli error on each flag qubit at each time step, propagate it, and record data-qubit errors, syndrome outcomes, and flag outcomes. Verify (1) that an X error on each flag produces exactly the data error shown in Figure 6, (2) that symmetric CNOTs cause cross-flag propagation to cancel so the originating flag is uniquely identified, and (3) that recomputing accepted-state logical error rates for the four initialization methods and the three codes at eta=0.5 and 100, with d1=3 extended to d2=9, reproduces Figures 8-10. If the propagated errors or flag outcomes differ, rerun the bias argument and the ordering.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline recommendation (ZXXZ-type XZZX code with down-triangle initialization) is derived from a mechanism that is asserted rather than demonstrated. Section III.A and Figure 5 claim that an X error on a flag qubit propagates through the CNOT chain to data qubits, that Z errors on flags propagate only to syndrome qubits, and that the symmetric CNOT layout makes propagated errors on other flag qubits cancel so that the originating flag is unambiguously identified. On this basis, Figure 6 concludes that the XZZX orientation injects extra X bias into data qubits while the ZXXZ orientation injects extra Z bias, and Section III.C argues that the ZXXZ orientation therefore reinforces the hardware's Z bias and wins. This is the load-bearing step: if the effective bias added by the flag-qubit network is not exactly as Figure 6 states, the XZZX/ZXXZ comparison changes and the recommended orientation could reverse. The manuscript provides neither the complete syndrome-extraction circuit nor the simulation code, so the error-propagation map cannot be independently checked from the text alone. There is also a tension with the post-selection protocol in Appendix A: if flag-triggered events are discarded in Stage I, it is not obvious how flag X errors can raise the accepted-state logical error rate, so the explanation may conflate raw data-qubit error rates with post-selected logical error rates.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript reports a numerical study of magic state injection on the heavy-hexagon architecture that uses flag qubits. It compares the surface code and the two orientations of the XZZX code (labeled XZZX-type and ZXXZ-type) under depolarizing and Z-biased noise, four data-qubit initialization schemes, and distance extension from d1=3 to d2 up to 9. Simulations are performed with Stim and decoded with minimum-weight perfect matching. The main claim is that flag qubits introduce an extra bias whose sign depends on the stabilizer orientation, so that under Z-biased noise the ZXXZ-type XZZX code with the down-triangle initialization method is the most suitable magic state injection choice for the heavy-hexagon structure.","tokens_in":33405,"tokens_out":7409,"duration_ms":68794,"significance":"If the central claim is correct, the paper provides practically useful guidance for magic state preparation on IBM-style heavy-hexagon hardware, where the combination of limited connectivity and biased noise makes a direct transfer of square-lattice injection protocols non-optimal. The paper's strengths are its use of standard simulation tools (Stim, MWPM), a clearly specified biased error model, and a falsifiable, concrete recommendation. The main weakness is that the load-bearing flag-qubit error-propagation mechanism is asserted through schematic diagrams rather than fully documented circuits, and the numerical evidence is presented without statistical uncertainties. With additional circuit-level details, post-selection-aware analysis, and confidence intervals, the recommendation would be substantially more convincing.","major_comments":[{"comment":"The paper explains the improved performance of the ZXXZ orientation by stating that X errors on flag qubits propagate to data qubits as extra X or Z errors and are identifiable through flag-qubit measurements. However, Appendix A describes a two-round post-selection procedure in which states with any detected error, including flag-triggered events, are discarded. If flag X errors are always identified and the corresponding states are discarded, they should not contribute to the logical error rate of accepted states. The manuscript needs to reconcile this tension: either the proposed extra-bias mechanism refers to raw, pre-post-selection error rates, or some flag X errors evade the flag measurement and survive post-selection. Please report accepted-state logical error rates together with flag-trigger statistics, and show explicitly how the extra-bias mechanism affects the post-selected logical error rate rather than only the rejected fraction of trials.","section":"Section III.A, Figs. 5-6 and Appendix A"},{"comment":"The claimed error-propagation map is not independently verifiable from the text. The paper states that X errors on a flag qubit propagate to data qubits, Z errors propagate to syndrome qubits, and the symmetric CNOT layout causes propagated errors on other flag qubits to cancel, leaving the originating flag error identifiable. Figure 5 shows only schematic fragments of the stabilizer measurement circuit. The authors should provide the complete syndrome-extraction circuit for one surface-code stabilizer and one XZZX/ZXXZ stabilizer, including all flag-qubit CNOTs and measurement/reset placements, or provide the actual Stim circuit used in the simulations. Without this, the extra-bias asymmetry in Figure 6 and the resulting recommendation cannot be checked by an independent reader.","section":"Section III.A and Fig. 5"},{"comment":"No error bars or confidence intervals are reported for any logical error rate. The paper draws conclusions from differences between initialization methods and between XZZX and ZXXZ orientations, some of which appear small relative to the reported scales. With 1e7 samples, binomial uncertainties are not negligible for logical error rates in the 1e-4 to 1e-3 range, especially when comparing closely spaced curves. Please add confidence intervals or statistical significance tests, and state the number of accepted post-selected samples used for each reported point, since post-selection can substantially reduce the effective sample size.","section":"Section III.B and Figs. 8-10"},{"comment":"The central explanation that the ZXXZ-type 'reinforces the Z bias' while the XZZX-type injects X bias remains qualitative. A quantitative validation would strengthen the claim: for example, estimate the effective bias per data qubit from the circuit-level propagation map or from simulation, and show that its sign and magnitude match the explanation; alternatively, vary the flag-qubit error rate independently of the data-qubit error rate and show that the XZZX/ZXXZ performance gap scales as predicted. Without such a test, the extra-bias mechanism is an interpretation of the observed trends rather than a demonstrated cause.","section":"Section III.C"}],"minor_comments":[{"comment":"The text says that a Z error in all data qubits in the 'first row' corresponds to a logical Z error, but Eq. (1) initializes a column of d-1 qubits. Given the logical Z operator defined in Section II.A as a top-to-bottom chain, the first column appears to be intended. Please correct the terminology.","section":"Section II.C, Eq. (1)"},{"comment":"The expression P_Z = 2ηP_single/(2(η+1)) has a redundant factor of 2; simplifying to ηP_single/(η+1) would avoid confusion.","section":"Section II.D, Eq. (4)"},{"comment":"The 'blind qubit' is not defined precisely. It is stated that one data qubit cannot have its errors detected, but it is not clear whether this is the magic-state qubit itself or a distinct boundary qubit, nor how its position is determined for each initialization method. Please provide an explicit definition.","section":"Section III.B and Fig. 7"},{"comment":"The physical error rate is said to include two-qubit gate error and readout error rates, but the error model in Section II.D does not specify how readout errors are modeled. Please clarify the readout error model and how it enters the simulation.","section":"Appendix A"},{"comment":"The phrase 'IBM's qubits typically exhibit T1 times longer than T2 times' is grammatically unclear; it should read that the T1 time is longer than the T2 time, or T1 >> T2.","section":"Introduction"}],"recommendation":"major_revision","confidential_remarks":"This is a simulation study squarely within the scope of a quantum error correction journal. The main risk is the load-bearing flag-qubit error-propagation mechanism and its interaction with post-selection; this is fixable with additional analysis, full circuits or code, and statistical details. The paper does not rely on circular reasoning, but it does depend on the authors' prior flag-qubit code construction; that dependency should be made explicit and the relevant circuits should be included. If the requested revisions are provided, the manuscript would likely be publishable, but in its current form the headline recommendation is not fully supported by the evidence presented."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent, applied simulation study of magic state injection on IBM-style heavy-hexagon hardware with flag qubits. The genuinely new part is the orientation-dependent bias effect: they split the XZZX code into XZZX-type and ZXXZ-type stabilizers, show that flag-qubit error propagation adds an effective X bias in one orientation and Z bias in the other, and recommend ZXXZ-type with down-triangle initialization under Z-biased noise. That recommendation is a plausible design rule for near-term FTQC, not a field-shifting result.\n\nWhat they do well: the error model (eqs 4-6) is consistent at the depolarizing limit, the simulation setup is standard (Stim + MWPM), and the comparison across bias, distance extension, and four initialization methods is thorough. They also correctly note the symmetry between XZZX-right and ZXXZ-down initializations, which is the right way to compare the two orientations.\n\nThe soft spots are real but not fatal. First, no error bars or confidence intervals on any logical error rate, and no code or data shipped; \"data appear in the article\" is not enough for a simulation paper. Second, and more load-bearing, the mechanism in Sec III.A / Figs 5-6 is asserted rather than demonstrated. The claim that X errors on flag qubits propagate to data qubits while Z errors go to syndrome qubits, and that symmetric CNOTs make propagated errors on other flags cancel, is the entire basis for the extra-bias story. If that error-propagation map is wrong, the XZZX/ZXXZ ordering could reverse. The full syndrome-extraction circuit is not given, so a referee cannot check this from the text. There is also an apparent tension with the post-selection protocol in Appendix A: if flag-triggered events are discarded in Stage I, how do flag X errors raise the accepted-state logical error rate? The paper does not resolve this; it may be conflating raw data-qubit error rates with post-selected logical error rates.\n\nThe citation pattern is fine; the main dependency is the authors' own prior flag-qubit code construction, cited as a given. That is a dependency, not circularity, but it means the result is only as solid as that prior construction.\n\nWho it's for: people working on QEC code adaptation to heavy-hex or low-degree hardware, and anyone planning magic state injection on IBM devices. It deserves a serious referee: the question is relevant, the study is honest, and the flaws are fixable. I would send it to review with the expectation that the authors provide code, error bars, and a clear accounting of how flag errors survive post-selection.","headline":"Competent applied simulation study; the orientation-dependent flag-qubit bias is plausible but the load-bearing error-propagation model is under-specified, so the recommendation should be treated as conditional until code, error bars, and post-selection logic are clarified.","tokens_in":33957,"tokens_out":3128,"would_cite":true,"duration_ms":26673,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68","81P70"],"pacs":["03.67.Pp","03.67.Lx"],"model":"deepseek-v4-flash","headline":"Heavy-hexagon magic states should be injected with the ZXXZ-oriented XZZX code and down-triangle initialization.","keywords":["magic state injection","heavy-hexagon architecture","XZZX code","flag qubits","biased noise","logical error rate","fault-tolerant quantum computing","quantum error correction"],"falsifier":"Run the stabilizer-measurement circuit of Figure 5(b) with a single $X$ error inserted on each flag qubit in turn and check whether the flag-qubit measurements identify the source while the syndrome measurements stay silent; if any single flag $X$ error produces an ambiguous or wrong signature, the extra-bias mechanism is not what the hardware does. Then compare the logical error rates of the XZZX and ZXXZ orientations on a real heavy-hexagon device under strong $Z$ bias; the paper's recommendation fails if the ZXXZ orientation does not come out lower.","tokens_in":32929,"feed_emoji":"⚛️","tokens_out":8105,"duration_ms":68867,"temperature":0.7,"pith_summary":"The paper aims to determine how to inject magic states—specially prepared qubit states that let a quantum computer run non-Clifford gates—on hardware whose qubits connect like a heavy-hexagon graph rather than a square lattice. Because such hardware needs flag qubits to measure stabilizers, the paper asks whether the usual XZZX code, its rotated ZXXZ form, or the surface code performs best under realistic biased noise. It claims the ZXXZ form combined with a triangular qubit-initialization layout is the most suitable choice, and explains the advantage by a flag-qubit error-propagation mechanism that adds an effective noise bias beyond the one put into the error model. If true, this gives a concrete recipe for fault-tolerant non-Clifford gates on near-term heavy-hexagon processors and shows that code orientation should be chosen with the flag-qubit wiring, not just the lattice, in mind.","feed_headline":"Flag qubits favor ZXXZ code for magic state injection","feed_subtitle":"On heavy-hexagon hardware, the ZXXZ orientation plus down-triangle initialization gives the lowest logical error rate.","key_machinery":"The load-bearing object is the flag-qubit error-propagation mechanism in the stabilizer measurement network: a CNOT sends $X$ errors from control to target and $Z$ errors from target to control, so an $X$ error on a flag qubit lands on data qubits while a $Z$ error lands only on syndrome readout, and the symmetric CNOT pattern cancels errors picked up by other flag qubits. This mechanism creates an extra per-data-qubit bias whose sign depends on the stabilizer orientation: the XZZX orientation adds $X$-type errors, while the ZXXZ orientation adds $Z$-type errors. The second mechanism is the initialization geometry: the four region divisions (right square, down square, right triangle, down triangle) determine which single 'blind' data qubit cannot be error-checked, and the down-triangle layout places the blind qubit where fewer flag-qubit errors reach it.","core_discovery":"The paper's central claim is that on a heavy-hexagon chip with flag qubits, the best way to inject a magic state is to encode it with the ZXXZ orientation of the XZZX code and initialize the surrounding data qubits with the down-triangle method. The reason is that flag qubits are not neutral: their $X$ errors propagate through the CNOT network into the data qubits, injecting an additional bias on top of the physical error model. The XZZX orientation adds extra $X$-type errors, weakening a $Z$ bias, while the ZXXZ orientation adds extra $Z$-type errors, reinforcing it; under the $Z$-biased noise typical of superconducting qubits, the ZXXZ orientation therefore has the lowest logical error rate. On a square lattice the two orientations are symmetric and equivalent, so the flag qubits are what break the symmetry.","pith_inferences":["A direct testable extension: if the hardware's dominant noise were $X$-biased rather than $Z$-biased, the same flag-qubit mechanism should make the XZZX orientation the better choice; the paper does not run this case.","The conclusion is specific to the symmetric CNOT layout: on other low-degree graphs where flag-qubit connections are asymmetric, the extra bias could point the other way or become position-dependent, so the orientation choice would need to be derived per layout.","Because the extra bias is generated by the stabilizer measurement circuit itself, the effective noise seen by the data qubits differs from the nominal physical error model; this suggests error-model calibration should measure effective per-qubit bias after wiring, not just gate errors."],"forward_implications":["On heavy-hexagon hardware with $Z$-biased noise, the ZXXZ orientation of the XZZX code should be used instead of the XZZX orientation for magic state injection.","Down-triangle initialization should be preferred at larger code distances, because triangle layouts keep error-detection regions compact and reduce the chain-error probability compared with square layouts.","Increasing the $Z$ bias lowers the heavy-hexagon logical error rate faster than on a square lattice because it suppresses the $X$ errors that flag qubits would otherwise inject into data qubits.","Errors introduced during the injection stage are not corrected by extending the code distance, so magic state distillation remains necessary even at low physical error rates.","The two XZZX orientations perform equivalently in the lattice structure, so any performance difference is a consequence of the flag-qubit wiring rather than of the code itself."],"supporting_citations":[{"why":"Supplies the magic state injection procedure on the rotated surface code that this work adapts to the heavy-hexagon layout.","marker":"[25]"},{"why":"Documents the heavy-hexagon connectivity and the $Z$-biased transmon hardware characteristics that motivate the comparison.","marker":"[31]"},{"why":"Introduces flag-qubit error correction on low-degree graphs, the mechanism whose error propagation is analyzed here.","marker":"[35]"},{"why":"Provides the prior construction of error-correcting codes for biased errors on the heavy-hexagon structure that this work extends.","marker":"[37]"},{"why":"Introduces the XZZX surface code whose two stabilizer orientations are compared.","marker":"[40]"},{"why":"Shows the XZZX code's advantage under biased noise, the baseline this paper builds on.","marker":"[41]"},{"why":"Provides the stabilizer-circuit simulator used to sample the magic state injection circuits.","marker":"[43]"},{"why":"Provides the decoder used to compute logical error rates from the syndrome samples.","marker":"[44]"}],"fun_headline_variants":["ZXXZ code favored on heavy-hexagon for magic states","Flag qubits tip magic injection to ZXXZ on heavy-hexagon","Heavy-hexagon magic states: ZXXZ with flags is key","ZXXZ beats XZZX for magic injection on heavy-hexagon"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's recommendation rests on the assumed error-propagation pattern through the flag-qubit network: an $X$ error on a flag qubit reaches the data, a $Z$ error reaches only the syndrome readout, and symmetric connections cancel secondary flag errors; if real hardware deviates from this pattern, the extra bias and the ZXXZ advantage disappear.","fun_headline_variants_meta":{"raw":{"variants":["ZXXZ code favored on heavy-hexagon for magic states","Flag qubits tip magic injection to ZXXZ on heavy-hexagon","Heavy-hexagon magic states: ZXXZ with flags is key","ZXXZ beats XZZX for magic injection on heavy-hexagon"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000333,"raw_usage":{"total_tokens":1859,"prompt_tokens":964,"completion_tokens":895,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":580,"completion_tokens_details":{"reasoning_tokens":814}},"tokens_in":580,"tokens_out":895,"duration_ms":7027,"temperature":1.0,"reasoning_tokens":814,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:07:07.661504+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the stabilizer-measurement circuit of Figure 5(b) with a single $X$ error inserted on each flag qubit in turn and check whether the flag-qubit measurements identify the source while the syndrome measurements stay silent; if any single flag $X$ error produces an ambiguous or wrong signature, the extra-bias mechanism is not what the hardware does. Then compare the logical error rates of the XZZX and ZXXZ orientations on a real heavy-hexagon device under strong $Z$ bias; the paper's recommendation fails if the ZXXZ orientation does not come out lower.","supporting_citations":[{"cited_title":"J., Srinivasan, S., Sundaresan, N., Bogorin, D","cited_arxiv_id":null,"evidence_quote":"Documents the heavy-hexagon connectivity and the $Z$-biased transmon hardware characteristics that motivate the comparison."},{"cited_title":"Magic state injection on the rotated surface code","cited_arxiv_id":null,"evidence_quote":"Supplies the magic state injection procedure on the rotated surface code that this work adapts to the heavy-hexagon layout."},{"cited_title":"J., Hertzberg, J","cited_arxiv_id":null,"evidence_quote":"Introduces flag-qubit error correction on low-degree graphs, the mechanism whose error propagation is analyzed here."},{"cited_title":"The XZZX surface code","cited_arxiv_id":null,"evidence_quote":"Introduces the XZZX surface code whose two stabilizer orientations are compared."},{"cited_title":"Practical quantum error correction with the XZZX code and Kerr-cat qubits","cited_arxiv_id":null,"evidence_quote":"Shows the XZZX code's advantage under biased noise, the baseline this paper builds on."},{"cited_title":"Stim: A fast stabilizer circuit simulator","cited_arxiv_id":null,"evidence_quote":"Provides the stabilizer-circuit simulator used to sample the magic state injection circuits."}],"review_version":1}