mathclaude-opus-4-20250514
Internal 5/5Mathematical 5/5
This work demonstrates exceptional mathematical rigor and internal consistency. The theoretical framework is clearly established with equation (1) providing the central scaling relationship εd ∝ (p/p_thr)^((d+1)/2) for below-threshold operation. All experimental results are analyzed within this framework consistently. The paper properly handles the subtleties of comparing physical and logical qubit lifetimes through appropriate channel fidelity metrics. Statistical uncertainties are propagated correctly through multi-level analysis. The mathematical validity is exemplary - all equations are dimensionally consistent, derivations proceed logically where shown, and approximations remain within stated validity regimes. While some standard results are stated without proof (average fidelity formulas, error budget decomposition), these are peripheral to the main mathematical argument and would be readily verifiable by specialists.
⚑Derivation Flags (23)
- medium
Main text / Supplement VI.B, ε_d = 1/2 (1 - (1 - 2p_L)^{1/t}) — Inference of logical error per cycle from a single t-cycle logical error probability is stated as a binomial/odd-parity formula, but the assumptions needed (effectively independent identical per-cycle logical flips or an equivalent Markov parity model) are not fully justified for correlated noise.If wrong: If the per-cycle logical process is not well captured by this parity model, then ε_d estimates in the one-point analyses of Fig. 2b and Fig. 3 could be biased, affecting quantitative scaling exponents and Λ extracted from those plots.
- medium
Main text Eq. (1) — Approximate surface-code scaling ε_d ∝ (p/p_thr)^{(d+1)/2} is presented without derivation in this manuscript and is used as the fitting/interpretive ansatz for Λ and threshold behavior.If wrong: If this scaling does not apply in the reported noise regime, then the interpretation of Λ as a threshold-style suppression factor and the comparison to ideal exponent (d+1)/2 would weaken, though the raw observed decrease of ε_d with d would still stand.
- medium
Main text Eq. (1): ε_d ∝ (p/p_thr)^{(d+1)/2} — Stated as an approximate scaling law for surface code logical error rate; no derivation or conditions (noise model class, decoder optimality, circuit-level vs phenomenological noise, finite-size corrections) are provided in-text.If wrong: If the exponent or functional form differs materially under the device’s correlated/circuit noise, then interpreting fitted slopes as a threshold signature and interpreting Λ as directly related to p_thr/p becomes less justified; the empirical observation that ε_d decreases with d remains valid.
- medium
Main text Sec. IV and SI Sec. VI(B): ε_d from a single point: ε_d = (1/2)(1-(1-2p_L)^{1/t}) — Inversion assumes independent identical per-cycle logical bit-flip on the relevant logical observable so that p_L(t) is the probability of an odd number of errors after t Bernoulli trials. Not derived; applicability under time-correlated bursts (explicitly present in the repetition-code data) is questionable.If wrong: Per-cycle ε_d inferred from long-t p_L could be biased (often downward in the presence of clustering/bursts), affecting reported floors such as ~10^{-10} per cycle and fitted Λ in regimes where rare events dominate.
- medium
Main text, break-even lifetime claim; Supplement IV.B eqs. (6)-(15) — The logical lifetime metric is constructed via average-fidelity-based effective depolarization rates and then bounded using Γ_logical ≤ 8ε_d/(3 t_c). This is mathematically plausible, but the argument depends on approximating the logical channel as a continuous Pauli channel with small error probabilities and on an unmeasured p_Y contribution.If wrong: If the logical noise is substantially non-Pauli or time-correlated, the quoted logical lifetime and gain factor 2.4±0.3 may not map cleanly onto the physical-qubit lifetime comparison, weakening the 'beyond break-even' quantification though not the raw memory data.
- medium
SI Sec. III(D) 'preweight correlations' and block/fuse streaming MWPM correctness — Algorithmic steps (conditioning on assumed edges; fusing blocks; declaring heralded failure if regions remain) are described heuristically without a proof that the resulting correction is equivalent to global MWPM on the full graph or without an error bound.If wrong: Could invalidate the claim that the streaming decoder returns a true global MWPM solution (and potentially bias ε_d for real-time decoding). The empirical accuracy comparisons in Fig. 4c would still stand, but theoretical claims about optimality/guarantees would be weakened.
- medium
SI Sec. IV(A) Eq. (5): (Λ_{3/5})^{-1} = Σ_i w_i p_expt^{(i)} — Linear decomposition of (Λ_{3/5})^{-1} into contributions from channels assumes (local) linear response and separability around an operating point; derivation of w_i and linearization regime are not shown here.If wrong: The numeric ‘error budget’ attribution fractions (e.g., 41% CZ, 20% idle, 17% correlated total) could be misleading; this would not negate observed Λ>2 but would weaken mechanistic conclusions about dominant contributors.
- medium
SI Sec. IV(A): 'Generalized Pauli Twirling Approximation' step — The twirling-to-generalized-Pauli channel step is asserted without a bound/justification for its effect on logical error rates in the presence of coherent errors/leakage/crosstalk; this is a modeling approximation.If wrong: Simulation agreement with experiment and extrapolations to larger distances (e.g., distance-27 projection) could be quantitatively unreliable; the experimental claims would remain intact.
- medium
SI Sec. IV(B) Eq. (13): Γ_logical ≤ 8 ε_d / (3 t_c) — Uses inequality p_X+p_Y+p_Z ≤ 2 ε_d (with ε_d defined as logical error probability per cycle averaged over X and Z eigenstates) and assumes the channel is well-approximated as Pauli with small probabilities. The link between ε_d (from odd-parity logical failure on X/Z eigenstates) and Pauli error rates is not fully derived.If wrong: The claimed strict upper bound and the asserted tightness (via p_Y suppression) may fail, affecting reported logical lifetime estimates and the break-even factor G; it would not affect the directly measured ε_d or Λ.
- medium
Supplement III.D.2, real-time decoder fusion/heralding construction — The block-fusion algorithm and the claim that M=10 yields negligible heralded failure relative to logical failure are described operationally but not proved or quantitatively bounded in the text.If wrong: If heralded failures are not negligible or the fusion approximation changes the effective decoding rule, the reported real-time decoding accuracy and latency tradeoff could be mischaracterized.
- medium
Supplement IV.A.2, extrapolation to d=27 and larger-distance forecasts — Projection to larger distances is based on fitted/simulated Λ treated as approximately stable with distance; detailed uncertainty or model-misspecification propagation is not shown.If wrong: If Λ degrades with distance or omitted correlated errors become dominant, resource estimates such as needing a distance-27 code for 10^-6 error could be substantially inaccurate; this affects outlook claims rather than the measured results.
- medium
Γ_logical = (4/3)(γ_X+γ_Y+γ_Z) and bound Γ_logical ≤ 8ε_d/(3 t_c) (Sec. IV.B) — The continuous-time channel approximation rests on small per-cycle error probabilities and on the assumption p_Y ≪ p_X, p_Z, justified only by reference to simulations (refs 37,38) without direct measurement of p_Y in this work.If wrong: If p_Y is not small, the equality 1/Γ_logical ≈ 291 μs underlying the break-even gain G = 2.4±0.3 could shift. Since the bound is strict, the gain remains at least lower-bounded, but the central quantitative claim of a 2.4× lifetime advantage could change in magnitude.
- low
Equation 5, Section IV.A.3 — Error budget decomposition (Λ3/5)^-1 = Σ wi p_expt^(i) stated without derivation of weight calculation methodologyIf wrong: Error budget percentages in Table S4 would be incorrect, but this is peripheral to main results
- low
Equations 7-9, Section IV.B — Average channel fidelity formulas for different noise models presented as known results without derivationIf wrong: Physical-logical lifetime comparison methodology would need revision, but main threshold results unaffected
- low
Error budget Eq. (5), Table S4 — Sensitivities w_i are extracted from simulations using a less-accurate correlated matching decoder, then linearly summed assuming independent contributions. Linear additivity at the operating point is stated but not proven; the simulation overpredicts Λ by 20%, which is acknowledged.If wrong: Quantitative attribution of error sources (e.g. the 60% CZ contribution) could shift, but the qualitative conclusion that CZ + idle errors dominate is robust to moderate deviations.
- low
Main text (definition/claim): Λ = ε_d/ε_{d+2} ≈ p_thr/p — The approximation Λ≈p_thr/p follows only if Eq. (1) holds with a distance-independent prefactor and if scaling is exactly geometric in (d+1)/2; neither is shown and later sections themselves note deviations (≈80% of ideal exponent).If wrong: Λ would remain a well-defined empirical suppression factor, but its interpretation as a direct estimator of p_thr/p would fail; conclusions about ‘below threshold’ based purely on Λ>1 would need more careful qualification (they do provide empirical crossings and simulations as additional context).
- low
One-point estimator ε_d = (1/2)[1-(1-2p_L)^(1/t)] (Sec. VI.B) — Derivation assumes independent identically distributed logical errors per cycle (binomial model). Time-correlated errors (e.g. leakage, the burst events explicitly observed in repetition codes) violate this assumption, but the estimator is still applied to those datasets including Fig. 3.If wrong: If correlations are significant, single-point ε_d estimates could systematically misrepresent per-cycle error; however, the paper's central Λ claims rely on multi-point exponential fits, not on this estimator, so the impact on main conclusions is limited.
- low
SI Sec. IV(A) Eqs. (3)-(4): Kraus operators for phenomenological DQLR reset model — K_0 = sqrt(I - Σ K_ij^† K_ij) is stated without discussing conditions ensuring the operator under the square root is positive semidefinite for the chosen transition probabilities P_{j→i}.If wrong: If parameters violate complete positivity, the simulated channel would be unphysical, undermining simulation-based comparisons/forecasts that rely on this DQLR model.
- low
SI Sec. IV(B) Eq. (10): short-time expansion F(δt)=1-(1/2)Γ δt — First-order (small-time) approximation is invoked; the range of δt for which it is accurate is not quantified relative to cycle time and extracted rates.If wrong: If higher-order terms are non-negligible over the times used, inferred Γ (and hence lifetimes) would be biased.
- low
SI Sec. IV(B) Eq. (12): Γ_logical = (4/3)(γ_X+γ_Y+γ_Z) — Derived from Eq. (9) under a continuous-time Pauli-rate model; not shown explicitly and relies on small-rate approximation and identification of per-cycle probabilities with rates via γ≈p/t_c.If wrong: Would affect conversion from per-cycle logical Pauli error probabilities to an effective depolarization lifetime; again mainly impacts the break-even lifetime-gain metric.
- low
SI Sec. IV(B) Eq. (7): F[E] = (1/12) Σ_{P=X,Y,Z} Tr[P E(P)] + 1/2 — Identity is quoted (from Nielsen 2002) but not derived; correct under trace-preserving channels with standard Pauli operator normalization conventions. The manuscript does not specify conventions, though the formula is standard.If wrong: If conventions differ, reported mapping from measured Pauli-eigenstate decays to average fidelity (and thus effective lifetime Γ^{-1}) could shift; relative comparisons might remain approximately correct if applied consistently.
- low
SI Sec. IV(B) Eqs. (8)-(9): average fidelity under amplitude damping/dephasing and Pauli channels — Formulas are stated without derivation; they depend on specific noise models (Markovian amplitude damping + dephasing; Pauli channel) and on interpreting the implemented process as that channel for duration t.If wrong: The effective-lifetime comparison between physical and logical qubits could be quantitatively off; however they also report empirical logical error per cycle, so only the '2.4× lifetime gain' metric would be affected.
- low
Supplement IV.A.3, eq. (5) error budget decomposition — The linear decomposition (Λ_{3/5})^{-1} = Σ_i w_i p_i is introduced as an error-budget model with sensitivities w_i, but the derivation of linearity around the operating point and extraction of w_i is only sketched.If wrong: If the budget is not approximately linear, the fractional attribution of contributions from CZ, idle, readout, and leakage errors would be unreliable, but the primary experimental demonstration of below-threshold scaling would remain unaffected.
+ Rigorous statistical uncertainty analysis with proper propagation through multi-level fits (Section VI)+ Clear mathematical framework distinguishing logical error per cycle εd from logical error probability pL with explicit conversion formula+ Proper treatment of finite-size effects and systematic corrections in threshold extrapolation
mathgpt-5.4-2026-03-05
Internal 4/5Mathematical 3/5
This submission is logically solid and mathematically competent, but not maximally rigorous in the theorem-proof sense. Its core empirical conclusion—that measured logical error per cycle decreases as code distance increases, with Λ > 2 over the tested range—is internally supported by the reported fitted error rates and does not appear self-contradictory. The notation and definitions are mostly stable, and the Supplement fills in several formulas that are only stated in the main text.
The main mathematical weakness is that a number of central quantitative interpretations depend on compressed modeling assumptions: the approximate threshold scaling law, the single-point conversion from p_L to ε_d, the effective-lifetime mapping used for break-even, the linearized error budget, and distance extrapolations. None of these looks obviously wrong, but several are load-bearing enough that they should be stated more explicitly as model-dependent rather than fully derived consequences. So the paper is mathematically credible, but some headline quantitative interpretations are less rigorously established than the raw experimental trends.
⚑Derivation Flags (23)
- medium
Main text / Supplement VI.B, ε_d = 1/2 (1 - (1 - 2p_L)^{1/t}) — Inference of logical error per cycle from a single t-cycle logical error probability is stated as a binomial/odd-parity formula, but the assumptions needed (effectively independent identical per-cycle logical flips or an equivalent Markov parity model) are not fully justified for correlated noise.If wrong: If the per-cycle logical process is not well captured by this parity model, then ε_d estimates in the one-point analyses of Fig. 2b and Fig. 3 could be biased, affecting quantitative scaling exponents and Λ extracted from those plots.
- medium
Main text Eq. (1) — Approximate surface-code scaling ε_d ∝ (p/p_thr)^{(d+1)/2} is presented without derivation in this manuscript and is used as the fitting/interpretive ansatz for Λ and threshold behavior.If wrong: If this scaling does not apply in the reported noise regime, then the interpretation of Λ as a threshold-style suppression factor and the comparison to ideal exponent (d+1)/2 would weaken, though the raw observed decrease of ε_d with d would still stand.
- medium
Main text Eq. (1): ε_d ∝ (p/p_thr)^{(d+1)/2} — Stated as an approximate scaling law for surface code logical error rate; no derivation or conditions (noise model class, decoder optimality, circuit-level vs phenomenological noise, finite-size corrections) are provided in-text.If wrong: If the exponent or functional form differs materially under the device’s correlated/circuit noise, then interpreting fitted slopes as a threshold signature and interpreting Λ as directly related to p_thr/p becomes less justified; the empirical observation that ε_d decreases with d remains valid.
- medium
Main text Sec. IV and SI Sec. VI(B): ε_d from a single point: ε_d = (1/2)(1-(1-2p_L)^{1/t}) — Inversion assumes independent identical per-cycle logical bit-flip on the relevant logical observable so that p_L(t) is the probability of an odd number of errors after t Bernoulli trials. Not derived; applicability under time-correlated bursts (explicitly present in the repetition-code data) is questionable.If wrong: Per-cycle ε_d inferred from long-t p_L could be biased (often downward in the presence of clustering/bursts), affecting reported floors such as ~10^{-10} per cycle and fitted Λ in regimes where rare events dominate.
- medium
Main text, break-even lifetime claim; Supplement IV.B eqs. (6)-(15) — The logical lifetime metric is constructed via average-fidelity-based effective depolarization rates and then bounded using Γ_logical ≤ 8ε_d/(3 t_c). This is mathematically plausible, but the argument depends on approximating the logical channel as a continuous Pauli channel with small error probabilities and on an unmeasured p_Y contribution.If wrong: If the logical noise is substantially non-Pauli or time-correlated, the quoted logical lifetime and gain factor 2.4±0.3 may not map cleanly onto the physical-qubit lifetime comparison, weakening the 'beyond break-even' quantification though not the raw memory data.
- medium
SI Sec. III(D) 'preweight correlations' and block/fuse streaming MWPM correctness — Algorithmic steps (conditioning on assumed edges; fusing blocks; declaring heralded failure if regions remain) are described heuristically without a proof that the resulting correction is equivalent to global MWPM on the full graph or without an error bound.If wrong: Could invalidate the claim that the streaming decoder returns a true global MWPM solution (and potentially bias ε_d for real-time decoding). The empirical accuracy comparisons in Fig. 4c would still stand, but theoretical claims about optimality/guarantees would be weakened.
- medium
SI Sec. IV(A) Eq. (5): (Λ_{3/5})^{-1} = Σ_i w_i p_expt^{(i)} — Linear decomposition of (Λ_{3/5})^{-1} into contributions from channels assumes (local) linear response and separability around an operating point; derivation of w_i and linearization regime are not shown here.If wrong: The numeric ‘error budget’ attribution fractions (e.g., 41% CZ, 20% idle, 17% correlated total) could be misleading; this would not negate observed Λ>2 but would weaken mechanistic conclusions about dominant contributors.
- medium
SI Sec. IV(A): 'Generalized Pauli Twirling Approximation' step — The twirling-to-generalized-Pauli channel step is asserted without a bound/justification for its effect on logical error rates in the presence of coherent errors/leakage/crosstalk; this is a modeling approximation.If wrong: Simulation agreement with experiment and extrapolations to larger distances (e.g., distance-27 projection) could be quantitatively unreliable; the experimental claims would remain intact.
- medium
SI Sec. IV(B) Eq. (13): Γ_logical ≤ 8 ε_d / (3 t_c) — Uses inequality p_X+p_Y+p_Z ≤ 2 ε_d (with ε_d defined as logical error probability per cycle averaged over X and Z eigenstates) and assumes the channel is well-approximated as Pauli with small probabilities. The link between ε_d (from odd-parity logical failure on X/Z eigenstates) and Pauli error rates is not fully derived.If wrong: The claimed strict upper bound and the asserted tightness (via p_Y suppression) may fail, affecting reported logical lifetime estimates and the break-even factor G; it would not affect the directly measured ε_d or Λ.
- medium
Supplement III.D.2, real-time decoder fusion/heralding construction — The block-fusion algorithm and the claim that M=10 yields negligible heralded failure relative to logical failure are described operationally but not proved or quantitatively bounded in the text.If wrong: If heralded failures are not negligible or the fusion approximation changes the effective decoding rule, the reported real-time decoding accuracy and latency tradeoff could be mischaracterized.
- medium
Supplement IV.A.2, extrapolation to d=27 and larger-distance forecasts — Projection to larger distances is based on fitted/simulated Λ treated as approximately stable with distance; detailed uncertainty or model-misspecification propagation is not shown.If wrong: If Λ degrades with distance or omitted correlated errors become dominant, resource estimates such as needing a distance-27 code for 10^-6 error could be substantially inaccurate; this affects outlook claims rather than the measured results.
- medium
Γ_logical = (4/3)(γ_X+γ_Y+γ_Z) and bound Γ_logical ≤ 8ε_d/(3 t_c) (Sec. IV.B) — The continuous-time channel approximation rests on small per-cycle error probabilities and on the assumption p_Y ≪ p_X, p_Z, justified only by reference to simulations (refs 37,38) without direct measurement of p_Y in this work.If wrong: If p_Y is not small, the equality 1/Γ_logical ≈ 291 μs underlying the break-even gain G = 2.4±0.3 could shift. Since the bound is strict, the gain remains at least lower-bounded, but the central quantitative claim of a 2.4× lifetime advantage could change in magnitude.
- low
Equation 5, Section IV.A.3 — Error budget decomposition (Λ3/5)^-1 = Σ wi p_expt^(i) stated without derivation of weight calculation methodologyIf wrong: Error budget percentages in Table S4 would be incorrect, but this is peripheral to main results
- low
Equations 7-9, Section IV.B — Average channel fidelity formulas for different noise models presented as known results without derivationIf wrong: Physical-logical lifetime comparison methodology would need revision, but main threshold results unaffected
- low
Error budget Eq. (5), Table S4 — Sensitivities w_i are extracted from simulations using a less-accurate correlated matching decoder, then linearly summed assuming independent contributions. Linear additivity at the operating point is stated but not proven; the simulation overpredicts Λ by 20%, which is acknowledged.If wrong: Quantitative attribution of error sources (e.g. the 60% CZ contribution) could shift, but the qualitative conclusion that CZ + idle errors dominate is robust to moderate deviations.
- low
Main text (definition/claim): Λ = ε_d/ε_{d+2} ≈ p_thr/p — The approximation Λ≈p_thr/p follows only if Eq. (1) holds with a distance-independent prefactor and if scaling is exactly geometric in (d+1)/2; neither is shown and later sections themselves note deviations (≈80% of ideal exponent).If wrong: Λ would remain a well-defined empirical suppression factor, but its interpretation as a direct estimator of p_thr/p would fail; conclusions about ‘below threshold’ based purely on Λ>1 would need more careful qualification (they do provide empirical crossings and simulations as additional context).
- low
One-point estimator ε_d = (1/2)[1-(1-2p_L)^(1/t)] (Sec. VI.B) — Derivation assumes independent identically distributed logical errors per cycle (binomial model). Time-correlated errors (e.g. leakage, the burst events explicitly observed in repetition codes) violate this assumption, but the estimator is still applied to those datasets including Fig. 3.If wrong: If correlations are significant, single-point ε_d estimates could systematically misrepresent per-cycle error; however, the paper's central Λ claims rely on multi-point exponential fits, not on this estimator, so the impact on main conclusions is limited.
- low
SI Sec. IV(A) Eqs. (3)-(4): Kraus operators for phenomenological DQLR reset model — K_0 = sqrt(I - Σ K_ij^† K_ij) is stated without discussing conditions ensuring the operator under the square root is positive semidefinite for the chosen transition probabilities P_{j→i}.If wrong: If parameters violate complete positivity, the simulated channel would be unphysical, undermining simulation-based comparisons/forecasts that rely on this DQLR model.
- low
SI Sec. IV(B) Eq. (10): short-time expansion F(δt)=1-(1/2)Γ δt — First-order (small-time) approximation is invoked; the range of δt for which it is accurate is not quantified relative to cycle time and extracted rates.If wrong: If higher-order terms are non-negligible over the times used, inferred Γ (and hence lifetimes) would be biased.
- low
SI Sec. IV(B) Eq. (12): Γ_logical = (4/3)(γ_X+γ_Y+γ_Z) — Derived from Eq. (9) under a continuous-time Pauli-rate model; not shown explicitly and relies on small-rate approximation and identification of per-cycle probabilities with rates via γ≈p/t_c.If wrong: Would affect conversion from per-cycle logical Pauli error probabilities to an effective depolarization lifetime; again mainly impacts the break-even lifetime-gain metric.
- low
SI Sec. IV(B) Eq. (7): F[E] = (1/12) Σ_{P=X,Y,Z} Tr[P E(P)] + 1/2 — Identity is quoted (from Nielsen 2002) but not derived; correct under trace-preserving channels with standard Pauli operator normalization conventions. The manuscript does not specify conventions, though the formula is standard.If wrong: If conventions differ, reported mapping from measured Pauli-eigenstate decays to average fidelity (and thus effective lifetime Γ^{-1}) could shift; relative comparisons might remain approximately correct if applied consistently.
- low
SI Sec. IV(B) Eqs. (8)-(9): average fidelity under amplitude damping/dephasing and Pauli channels — Formulas are stated without derivation; they depend on specific noise models (Markovian amplitude damping + dephasing; Pauli channel) and on interpreting the implemented process as that channel for duration t.If wrong: The effective-lifetime comparison between physical and logical qubits could be quantitatively off; however they also report empirical logical error per cycle, so only the '2.4× lifetime gain' metric would be affected.
- low
Supplement IV.A.3, eq. (5) error budget decomposition — The linear decomposition (Λ_{3/5})^{-1} = Σ_i w_i p_i is introduced as an error-budget model with sensitivities w_i, but the derivation of linearity around the operating point and extraction of w_i is only sketched.If wrong: If the budget is not approximately linear, the fractional attribution of contributions from CZ, idle, readout, and leakage errors would be unreliable, but the primary experimental demonstration of below-threshold scaling would remain unaffected.
+ The relation between cumulative logical error probability p_L and per-cycle error ε_d is explicitly stated and used consistently; the parity formula in Supplement VI.B is mathematically coherent within its assumed model.+ Dimensional consistency is good in the lifetime analysis: formulas involving T1, T2, Γ, ε_d, and cycle time t_c in Supplement IV.B have correct units and a clear chain from average fidelity to effective lifetime.+ The paper distinguishes direct measurements from simulation/model-based inferences reasonably well, e.g. experimental ε_d and Λ versus simulated larger-distance projections and error-budget decomposition.
- Main text Eq. (1) is an approximate ansatz, but later interpretations of Λ and resource extrapolations rely on it more strongly than the manuscript mathematically justifies.- The one-point inference ε_d = 1/2(1-(1-2p_L)^{1/t}) assumes a specific per-cycle logical error model; this assumption is not reconciled with the correlated burst events that dominate the low-error repetition-code regime.- The break-even lifetime claim depends on modeling the logical channel as an effective continuous Pauli channel and on bounding unmeasured components such as p_Y; the resulting 2.4±0.3 gain is therefore somewhat model-dependent.- The error-budget decomposition in Supplement IV.A.3 is presented as linear in component error probabilities around the operating point, but the validity range of this linearization and extraction of sensitivities w_i is not shown.- Real-time decoder claims rely on an algorithmic construction whose correctness/performance bounds are described procedurally rather than proved; the statement that heralded failure is negligible is asserted without a quantitative bound in the text.
mathclaude-opus-4-7
Internal 5/5Mathematical 5/5
The submission is mathematically and logically rigorous within its empirical scope. Definitions are stable, fits are appropriate, uncertainty propagation is documented in the supplementary material, and approximations are flagged with their validity regimes. The central claims — Λ = 2.14±0.02, ε_7 = 0.143%, and a 2.4× lifetime gain over the best physical qubit — follow cleanly from the stated procedures and are internally consistent across main text, tables, and figures. Importantly, the break-even claim is anchored by a strict mathematical upper bound rather than an approximation treated as exact, and the multiple independent decoders (neural network, ensembled matching, real-time) yield consistent orderings of ε_d that support the robustness of the conclusions. The minor concerns identified (i.i.d. assumption in the one-point estimator, p_Y suppression assumption in the lifetime conversion, 20% unmodeled error in the budget) are acknowledged in the paper and do not undermine the central conclusions; they affect quantitative magnitudes at the 10-20% level rather than the qualitative below-threshold claim. No circular derivations, definition drifts, or undeclared approximation escalations were identified.
⚑Derivation Flags (23)
- medium
Main text / Supplement VI.B, ε_d = 1/2 (1 - (1 - 2p_L)^{1/t}) — Inference of logical error per cycle from a single t-cycle logical error probability is stated as a binomial/odd-parity formula, but the assumptions needed (effectively independent identical per-cycle logical flips or an equivalent Markov parity model) are not fully justified for correlated noise.If wrong: If the per-cycle logical process is not well captured by this parity model, then ε_d estimates in the one-point analyses of Fig. 2b and Fig. 3 could be biased, affecting quantitative scaling exponents and Λ extracted from those plots.
- medium
Main text Eq. (1) — Approximate surface-code scaling ε_d ∝ (p/p_thr)^{(d+1)/2} is presented without derivation in this manuscript and is used as the fitting/interpretive ansatz for Λ and threshold behavior.If wrong: If this scaling does not apply in the reported noise regime, then the interpretation of Λ as a threshold-style suppression factor and the comparison to ideal exponent (d+1)/2 would weaken, though the raw observed decrease of ε_d with d would still stand.
- medium
Main text Eq. (1): ε_d ∝ (p/p_thr)^{(d+1)/2} — Stated as an approximate scaling law for surface code logical error rate; no derivation or conditions (noise model class, decoder optimality, circuit-level vs phenomenological noise, finite-size corrections) are provided in-text.If wrong: If the exponent or functional form differs materially under the device’s correlated/circuit noise, then interpreting fitted slopes as a threshold signature and interpreting Λ as directly related to p_thr/p becomes less justified; the empirical observation that ε_d decreases with d remains valid.
- medium
Main text Sec. IV and SI Sec. VI(B): ε_d from a single point: ε_d = (1/2)(1-(1-2p_L)^{1/t}) — Inversion assumes independent identical per-cycle logical bit-flip on the relevant logical observable so that p_L(t) is the probability of an odd number of errors after t Bernoulli trials. Not derived; applicability under time-correlated bursts (explicitly present in the repetition-code data) is questionable.If wrong: Per-cycle ε_d inferred from long-t p_L could be biased (often downward in the presence of clustering/bursts), affecting reported floors such as ~10^{-10} per cycle and fitted Λ in regimes where rare events dominate.
- medium
Main text, break-even lifetime claim; Supplement IV.B eqs. (6)-(15) — The logical lifetime metric is constructed via average-fidelity-based effective depolarization rates and then bounded using Γ_logical ≤ 8ε_d/(3 t_c). This is mathematically plausible, but the argument depends on approximating the logical channel as a continuous Pauli channel with small error probabilities and on an unmeasured p_Y contribution.If wrong: If the logical noise is substantially non-Pauli or time-correlated, the quoted logical lifetime and gain factor 2.4±0.3 may not map cleanly onto the physical-qubit lifetime comparison, weakening the 'beyond break-even' quantification though not the raw memory data.
- medium
SI Sec. III(D) 'preweight correlations' and block/fuse streaming MWPM correctness — Algorithmic steps (conditioning on assumed edges; fusing blocks; declaring heralded failure if regions remain) are described heuristically without a proof that the resulting correction is equivalent to global MWPM on the full graph or without an error bound.If wrong: Could invalidate the claim that the streaming decoder returns a true global MWPM solution (and potentially bias ε_d for real-time decoding). The empirical accuracy comparisons in Fig. 4c would still stand, but theoretical claims about optimality/guarantees would be weakened.
- medium
SI Sec. IV(A) Eq. (5): (Λ_{3/5})^{-1} = Σ_i w_i p_expt^{(i)} — Linear decomposition of (Λ_{3/5})^{-1} into contributions from channels assumes (local) linear response and separability around an operating point; derivation of w_i and linearization regime are not shown here.If wrong: The numeric ‘error budget’ attribution fractions (e.g., 41% CZ, 20% idle, 17% correlated total) could be misleading; this would not negate observed Λ>2 but would weaken mechanistic conclusions about dominant contributors.
- medium
SI Sec. IV(A): 'Generalized Pauli Twirling Approximation' step — The twirling-to-generalized-Pauli channel step is asserted without a bound/justification for its effect on logical error rates in the presence of coherent errors/leakage/crosstalk; this is a modeling approximation.If wrong: Simulation agreement with experiment and extrapolations to larger distances (e.g., distance-27 projection) could be quantitatively unreliable; the experimental claims would remain intact.
- medium
SI Sec. IV(B) Eq. (13): Γ_logical ≤ 8 ε_d / (3 t_c) — Uses inequality p_X+p_Y+p_Z ≤ 2 ε_d (with ε_d defined as logical error probability per cycle averaged over X and Z eigenstates) and assumes the channel is well-approximated as Pauli with small probabilities. The link between ε_d (from odd-parity logical failure on X/Z eigenstates) and Pauli error rates is not fully derived.If wrong: The claimed strict upper bound and the asserted tightness (via p_Y suppression) may fail, affecting reported logical lifetime estimates and the break-even factor G; it would not affect the directly measured ε_d or Λ.
- medium
Supplement III.D.2, real-time decoder fusion/heralding construction — The block-fusion algorithm and the claim that M=10 yields negligible heralded failure relative to logical failure are described operationally but not proved or quantitatively bounded in the text.If wrong: If heralded failures are not negligible or the fusion approximation changes the effective decoding rule, the reported real-time decoding accuracy and latency tradeoff could be mischaracterized.
- medium
Supplement IV.A.2, extrapolation to d=27 and larger-distance forecasts — Projection to larger distances is based on fitted/simulated Λ treated as approximately stable with distance; detailed uncertainty or model-misspecification propagation is not shown.If wrong: If Λ degrades with distance or omitted correlated errors become dominant, resource estimates such as needing a distance-27 code for 10^-6 error could be substantially inaccurate; this affects outlook claims rather than the measured results.
- medium
Γ_logical = (4/3)(γ_X+γ_Y+γ_Z) and bound Γ_logical ≤ 8ε_d/(3 t_c) (Sec. IV.B) — The continuous-time channel approximation rests on small per-cycle error probabilities and on the assumption p_Y ≪ p_X, p_Z, justified only by reference to simulations (refs 37,38) without direct measurement of p_Y in this work.If wrong: If p_Y is not small, the equality 1/Γ_logical ≈ 291 μs underlying the break-even gain G = 2.4±0.3 could shift. Since the bound is strict, the gain remains at least lower-bounded, but the central quantitative claim of a 2.4× lifetime advantage could change in magnitude.
- low
Equation 5, Section IV.A.3 — Error budget decomposition (Λ3/5)^-1 = Σ wi p_expt^(i) stated without derivation of weight calculation methodologyIf wrong: Error budget percentages in Table S4 would be incorrect, but this is peripheral to main results
- low
Equations 7-9, Section IV.B — Average channel fidelity formulas for different noise models presented as known results without derivationIf wrong: Physical-logical lifetime comparison methodology would need revision, but main threshold results unaffected
- low
Error budget Eq. (5), Table S4 — Sensitivities w_i are extracted from simulations using a less-accurate correlated matching decoder, then linearly summed assuming independent contributions. Linear additivity at the operating point is stated but not proven; the simulation overpredicts Λ by 20%, which is acknowledged.If wrong: Quantitative attribution of error sources (e.g. the 60% CZ contribution) could shift, but the qualitative conclusion that CZ + idle errors dominate is robust to moderate deviations.
- low
Main text (definition/claim): Λ = ε_d/ε_{d+2} ≈ p_thr/p — The approximation Λ≈p_thr/p follows only if Eq. (1) holds with a distance-independent prefactor and if scaling is exactly geometric in (d+1)/2; neither is shown and later sections themselves note deviations (≈80% of ideal exponent).If wrong: Λ would remain a well-defined empirical suppression factor, but its interpretation as a direct estimator of p_thr/p would fail; conclusions about ‘below threshold’ based purely on Λ>1 would need more careful qualification (they do provide empirical crossings and simulations as additional context).
- low
One-point estimator ε_d = (1/2)[1-(1-2p_L)^(1/t)] (Sec. VI.B) — Derivation assumes independent identically distributed logical errors per cycle (binomial model). Time-correlated errors (e.g. leakage, the burst events explicitly observed in repetition codes) violate this assumption, but the estimator is still applied to those datasets including Fig. 3.If wrong: If correlations are significant, single-point ε_d estimates could systematically misrepresent per-cycle error; however, the paper's central Λ claims rely on multi-point exponential fits, not on this estimator, so the impact on main conclusions is limited.
- low
SI Sec. IV(A) Eqs. (3)-(4): Kraus operators for phenomenological DQLR reset model — K_0 = sqrt(I - Σ K_ij^† K_ij) is stated without discussing conditions ensuring the operator under the square root is positive semidefinite for the chosen transition probabilities P_{j→i}.If wrong: If parameters violate complete positivity, the simulated channel would be unphysical, undermining simulation-based comparisons/forecasts that rely on this DQLR model.
- low
SI Sec. IV(B) Eq. (10): short-time expansion F(δt)=1-(1/2)Γ δt — First-order (small-time) approximation is invoked; the range of δt for which it is accurate is not quantified relative to cycle time and extracted rates.If wrong: If higher-order terms are non-negligible over the times used, inferred Γ (and hence lifetimes) would be biased.
- low
SI Sec. IV(B) Eq. (12): Γ_logical = (4/3)(γ_X+γ_Y+γ_Z) — Derived from Eq. (9) under a continuous-time Pauli-rate model; not shown explicitly and relies on small-rate approximation and identification of per-cycle probabilities with rates via γ≈p/t_c.If wrong: Would affect conversion from per-cycle logical Pauli error probabilities to an effective depolarization lifetime; again mainly impacts the break-even lifetime-gain metric.
- low
SI Sec. IV(B) Eq. (7): F[E] = (1/12) Σ_{P=X,Y,Z} Tr[P E(P)] + 1/2 — Identity is quoted (from Nielsen 2002) but not derived; correct under trace-preserving channels with standard Pauli operator normalization conventions. The manuscript does not specify conventions, though the formula is standard.If wrong: If conventions differ, reported mapping from measured Pauli-eigenstate decays to average fidelity (and thus effective lifetime Γ^{-1}) could shift; relative comparisons might remain approximately correct if applied consistently.
- low
SI Sec. IV(B) Eqs. (8)-(9): average fidelity under amplitude damping/dephasing and Pauli channels — Formulas are stated without derivation; they depend on specific noise models (Markovian amplitude damping + dephasing; Pauli channel) and on interpreting the implemented process as that channel for duration t.If wrong: The effective-lifetime comparison between physical and logical qubits could be quantitatively off; however they also report empirical logical error per cycle, so only the '2.4× lifetime gain' metric would be affected.
- low
Supplement IV.A.3, eq. (5) error budget decomposition — The linear decomposition (Λ_{3/5})^{-1} = Σ_i w_i p_i is introduced as an error-budget model with sensitivities w_i, but the derivation of linearity around the operating point and extraction of w_i is only sketched.If wrong: If the budget is not approximately linear, the fractional attribution of contributions from CZ, idle, readout, and leakage errors would be unreliable, but the primary experimental demonstration of below-threshold scaling would remain unaffected.
+ Quantitative claims are supported by explicit fitting procedures with uncertainties propagated from both statistical sampling and drift residuals; tables (S1, S5) give per-grid, per-basis breakdowns enabling reproducibility.+ The break-even claim is grounded in a rigorously defined, basis-uniform lifetime metric (Eqs. 6-13 of SI) with a strict upper bound Γ_logical ≤ 8ε_d/(3t_c) rather than an unjustified equality.+ The scaling-law test in Eq. 1 is probed both by surface codes (d=3,5,7) and by repetition codes up to d=29 with coherent error injection, providing strong internal cross-checks of the (d+1)/2 exponent prediction.
- The continuous-time logical channel approximation underlying 1/Γ_logical = 291 μs assumes p_Y ≪ p_X, p_Z, which is justified only by reference to prior simulations rather than direct measurement on this device; the central gain factor 2.4 depends on this assumption (mitigated by the strict upper bound).- The one-point estimator ε_d = (1/2)[1-(1-2p_L)^(1/t)] assumes i.i.d. per-cycle errors; this is applied to the d=29 repetition-code data where rare correlated bursts are explicitly identified as dominant, so the inferred 10^-10 floor mixes a binomial estimator with non-binomial physics. The paper acknowledges the floor is inferred rather than measured (no errors observed at d=29).- The error budget in Table S4 reports that simulation overpredicts Λ by 20%, indicating ~20% of error mechanisms are unmodeled; the attribution percentages (e.g. 60% from CZ) should be read as approximate rather than tight.- Reported deviation of measured exponent from the ideal (d+1)/2 (~80% of ideal) is attributed to 'excess correlations' without a quantitative model, leaving a small but unmodeled gap between theory and observation.
mathgpt-5.2-2025-12-11
Internal 4/5Mathematical 3/5
Mathematically, the submission is largely an experimental report with a small number of explicit formulas. The formulas that are explicitly manipulated (especially the odd-parity logical-failure accumulation and its inversion, and the regression-based extraction of Λ) are algebraically correct and used consistently. There are no evident self-contradictions in definitions of ε_d, p_L, Λ, p_det, or the decoding/latency metrics.
Where rigor weakens is in several modeling-to-interpretation bridges that are asserted rather than derived: the canonical surface-code scaling law Eq. (1) as applied to this circuit-level, correlated-noise setting; the approximation Λ≈p_thr/p; the first-order linear error-budget decomposition; and the simulation approximations (Pauli twirling/leakage handling) used for extrapolation. Additionally, the per-cycle logical error inferred from a single long-duration point can be systematically distorted in the presence of the very rare correlated bursts the paper identifies. These issues primarily affect interpretation, attribution, and extrapolation—not the internal coherence of the reported empirical measurements.
⚑Derivation Flags (23)
- medium
Main text / Supplement VI.B, ε_d = 1/2 (1 - (1 - 2p_L)^{1/t}) — Inference of logical error per cycle from a single t-cycle logical error probability is stated as a binomial/odd-parity formula, but the assumptions needed (effectively independent identical per-cycle logical flips or an equivalent Markov parity model) are not fully justified for correlated noise.If wrong: If the per-cycle logical process is not well captured by this parity model, then ε_d estimates in the one-point analyses of Fig. 2b and Fig. 3 could be biased, affecting quantitative scaling exponents and Λ extracted from those plots.
- medium
Main text Eq. (1) — Approximate surface-code scaling ε_d ∝ (p/p_thr)^{(d+1)/2} is presented without derivation in this manuscript and is used as the fitting/interpretive ansatz for Λ and threshold behavior.If wrong: If this scaling does not apply in the reported noise regime, then the interpretation of Λ as a threshold-style suppression factor and the comparison to ideal exponent (d+1)/2 would weaken, though the raw observed decrease of ε_d with d would still stand.
- medium
Main text Eq. (1): ε_d ∝ (p/p_thr)^{(d+1)/2} — Stated as an approximate scaling law for surface code logical error rate; no derivation or conditions (noise model class, decoder optimality, circuit-level vs phenomenological noise, finite-size corrections) are provided in-text.If wrong: If the exponent or functional form differs materially under the device’s correlated/circuit noise, then interpreting fitted slopes as a threshold signature and interpreting Λ as directly related to p_thr/p becomes less justified; the empirical observation that ε_d decreases with d remains valid.
- medium
Main text Sec. IV and SI Sec. VI(B): ε_d from a single point: ε_d = (1/2)(1-(1-2p_L)^{1/t}) — Inversion assumes independent identical per-cycle logical bit-flip on the relevant logical observable so that p_L(t) is the probability of an odd number of errors after t Bernoulli trials. Not derived; applicability under time-correlated bursts (explicitly present in the repetition-code data) is questionable.If wrong: Per-cycle ε_d inferred from long-t p_L could be biased (often downward in the presence of clustering/bursts), affecting reported floors such as ~10^{-10} per cycle and fitted Λ in regimes where rare events dominate.
- medium
Main text, break-even lifetime claim; Supplement IV.B eqs. (6)-(15) — The logical lifetime metric is constructed via average-fidelity-based effective depolarization rates and then bounded using Γ_logical ≤ 8ε_d/(3 t_c). This is mathematically plausible, but the argument depends on approximating the logical channel as a continuous Pauli channel with small error probabilities and on an unmeasured p_Y contribution.If wrong: If the logical noise is substantially non-Pauli or time-correlated, the quoted logical lifetime and gain factor 2.4±0.3 may not map cleanly onto the physical-qubit lifetime comparison, weakening the 'beyond break-even' quantification though not the raw memory data.
- medium
SI Sec. III(D) 'preweight correlations' and block/fuse streaming MWPM correctness — Algorithmic steps (conditioning on assumed edges; fusing blocks; declaring heralded failure if regions remain) are described heuristically without a proof that the resulting correction is equivalent to global MWPM on the full graph or without an error bound.If wrong: Could invalidate the claim that the streaming decoder returns a true global MWPM solution (and potentially bias ε_d for real-time decoding). The empirical accuracy comparisons in Fig. 4c would still stand, but theoretical claims about optimality/guarantees would be weakened.
- medium
SI Sec. IV(A) Eq. (5): (Λ_{3/5})^{-1} = Σ_i w_i p_expt^{(i)} — Linear decomposition of (Λ_{3/5})^{-1} into contributions from channels assumes (local) linear response and separability around an operating point; derivation of w_i and linearization regime are not shown here.If wrong: The numeric ‘error budget’ attribution fractions (e.g., 41% CZ, 20% idle, 17% correlated total) could be misleading; this would not negate observed Λ>2 but would weaken mechanistic conclusions about dominant contributors.
- medium
SI Sec. IV(A): 'Generalized Pauli Twirling Approximation' step — The twirling-to-generalized-Pauli channel step is asserted without a bound/justification for its effect on logical error rates in the presence of coherent errors/leakage/crosstalk; this is a modeling approximation.If wrong: Simulation agreement with experiment and extrapolations to larger distances (e.g., distance-27 projection) could be quantitatively unreliable; the experimental claims would remain intact.
- medium
SI Sec. IV(B) Eq. (13): Γ_logical ≤ 8 ε_d / (3 t_c) — Uses inequality p_X+p_Y+p_Z ≤ 2 ε_d (with ε_d defined as logical error probability per cycle averaged over X and Z eigenstates) and assumes the channel is well-approximated as Pauli with small probabilities. The link between ε_d (from odd-parity logical failure on X/Z eigenstates) and Pauli error rates is not fully derived.If wrong: The claimed strict upper bound and the asserted tightness (via p_Y suppression) may fail, affecting reported logical lifetime estimates and the break-even factor G; it would not affect the directly measured ε_d or Λ.
- medium
Supplement III.D.2, real-time decoder fusion/heralding construction — The block-fusion algorithm and the claim that M=10 yields negligible heralded failure relative to logical failure are described operationally but not proved or quantitatively bounded in the text.If wrong: If heralded failures are not negligible or the fusion approximation changes the effective decoding rule, the reported real-time decoding accuracy and latency tradeoff could be mischaracterized.
- medium
Supplement IV.A.2, extrapolation to d=27 and larger-distance forecasts — Projection to larger distances is based on fitted/simulated Λ treated as approximately stable with distance; detailed uncertainty or model-misspecification propagation is not shown.If wrong: If Λ degrades with distance or omitted correlated errors become dominant, resource estimates such as needing a distance-27 code for 10^-6 error could be substantially inaccurate; this affects outlook claims rather than the measured results.
- medium
Γ_logical = (4/3)(γ_X+γ_Y+γ_Z) and bound Γ_logical ≤ 8ε_d/(3 t_c) (Sec. IV.B) — The continuous-time channel approximation rests on small per-cycle error probabilities and on the assumption p_Y ≪ p_X, p_Z, justified only by reference to simulations (refs 37,38) without direct measurement of p_Y in this work.If wrong: If p_Y is not small, the equality 1/Γ_logical ≈ 291 μs underlying the break-even gain G = 2.4±0.3 could shift. Since the bound is strict, the gain remains at least lower-bounded, but the central quantitative claim of a 2.4× lifetime advantage could change in magnitude.
- low
Equation 5, Section IV.A.3 — Error budget decomposition (Λ3/5)^-1 = Σ wi p_expt^(i) stated without derivation of weight calculation methodologyIf wrong: Error budget percentages in Table S4 would be incorrect, but this is peripheral to main results
- low
Equations 7-9, Section IV.B — Average channel fidelity formulas for different noise models presented as known results without derivationIf wrong: Physical-logical lifetime comparison methodology would need revision, but main threshold results unaffected
- low
Error budget Eq. (5), Table S4 — Sensitivities w_i are extracted from simulations using a less-accurate correlated matching decoder, then linearly summed assuming independent contributions. Linear additivity at the operating point is stated but not proven; the simulation overpredicts Λ by 20%, which is acknowledged.If wrong: Quantitative attribution of error sources (e.g. the 60% CZ contribution) could shift, but the qualitative conclusion that CZ + idle errors dominate is robust to moderate deviations.
- low
Main text (definition/claim): Λ = ε_d/ε_{d+2} ≈ p_thr/p — The approximation Λ≈p_thr/p follows only if Eq. (1) holds with a distance-independent prefactor and if scaling is exactly geometric in (d+1)/2; neither is shown and later sections themselves note deviations (≈80% of ideal exponent).If wrong: Λ would remain a well-defined empirical suppression factor, but its interpretation as a direct estimator of p_thr/p would fail; conclusions about ‘below threshold’ based purely on Λ>1 would need more careful qualification (they do provide empirical crossings and simulations as additional context).
- low
One-point estimator ε_d = (1/2)[1-(1-2p_L)^(1/t)] (Sec. VI.B) — Derivation assumes independent identically distributed logical errors per cycle (binomial model). Time-correlated errors (e.g. leakage, the burst events explicitly observed in repetition codes) violate this assumption, but the estimator is still applied to those datasets including Fig. 3.If wrong: If correlations are significant, single-point ε_d estimates could systematically misrepresent per-cycle error; however, the paper's central Λ claims rely on multi-point exponential fits, not on this estimator, so the impact on main conclusions is limited.
- low
SI Sec. IV(A) Eqs. (3)-(4): Kraus operators for phenomenological DQLR reset model — K_0 = sqrt(I - Σ K_ij^† K_ij) is stated without discussing conditions ensuring the operator under the square root is positive semidefinite for the chosen transition probabilities P_{j→i}.If wrong: If parameters violate complete positivity, the simulated channel would be unphysical, undermining simulation-based comparisons/forecasts that rely on this DQLR model.
- low
SI Sec. IV(B) Eq. (10): short-time expansion F(δt)=1-(1/2)Γ δt — First-order (small-time) approximation is invoked; the range of δt for which it is accurate is not quantified relative to cycle time and extracted rates.If wrong: If higher-order terms are non-negligible over the times used, inferred Γ (and hence lifetimes) would be biased.
- low
SI Sec. IV(B) Eq. (12): Γ_logical = (4/3)(γ_X+γ_Y+γ_Z) — Derived from Eq. (9) under a continuous-time Pauli-rate model; not shown explicitly and relies on small-rate approximation and identification of per-cycle probabilities with rates via γ≈p/t_c.If wrong: Would affect conversion from per-cycle logical Pauli error probabilities to an effective depolarization lifetime; again mainly impacts the break-even lifetime-gain metric.
- low
SI Sec. IV(B) Eq. (7): F[E] = (1/12) Σ_{P=X,Y,Z} Tr[P E(P)] + 1/2 — Identity is quoted (from Nielsen 2002) but not derived; correct under trace-preserving channels with standard Pauli operator normalization conventions. The manuscript does not specify conventions, though the formula is standard.If wrong: If conventions differ, reported mapping from measured Pauli-eigenstate decays to average fidelity (and thus effective lifetime Γ^{-1}) could shift; relative comparisons might remain approximately correct if applied consistently.
- low
SI Sec. IV(B) Eqs. (8)-(9): average fidelity under amplitude damping/dephasing and Pauli channels — Formulas are stated without derivation; they depend on specific noise models (Markovian amplitude damping + dephasing; Pauli channel) and on interpreting the implemented process as that channel for duration t.If wrong: The effective-lifetime comparison between physical and logical qubits could be quantitatively off; however they also report empirical logical error per cycle, so only the '2.4× lifetime gain' metric would be affected.
- low
Supplement IV.A.3, eq. (5) error budget decomposition — The linear decomposition (Λ_{3/5})^{-1} = Σ_i w_i p_i is introduced as an error-budget model with sensitivities w_i, but the derivation of linearity around the operating point and extraction of w_i is only sketched.If wrong: If the budget is not approximately linear, the fractional attribution of contributions from CZ, idle, readout, and leakage errors would be unreliable, but the primary experimental demonstration of below-threshold scaling would remain unaffected.
+ Correct and consistent use of the standard odd-parity accumulation formula for logical error probability under an i.i.d. per-cycle logical-flip model, including explicit inversion to ε_d (SI VI(B)).+ Uncertainty propagation for Λ via regression on ln ε_d vs (d+1)/2 is internally coherent and appropriately treats Λ as a fitted parameter rather than an identity (SI VI(A)).+ The lifetime comparison framework is at least mathematically structured: define average fidelity (Eq. (6)), relate to Pauli observables (Eq. (7)), define effective rate Γ via short-time expansion (Eq. (10)), then compare 1/Γ across physical/logical (SI IV(B)).
- Eq. (1) scaling and the derived heuristic Λ≈p_thr/p are not derived nor are their validity conditions stated; later quantitative interpretations lean on them (below-threshold narrative; interpreting detection probability as proxy for 1/Λ).- The one-point ε_d inference from p_L at t=1000 is mathematically correct only under an i.i.d. per-cycle logical error model; the same section explicitly identifies rare, temporally correlated burst events that can violate this assumption, potentially biasing ε_d and the claimed ~10^{-10} floor.- Simulation pipeline includes a generalized Pauli twirling approximation and phenomenological leakage-reset modeling; without bounds, extrapolations to larger distances (e.g., distance-27 projection) and mechanistic claims from simulation agreement are mathematically under-justified.- The linear ‘1/Λ error budget’ decomposition (Λ^{-1} = Σ w_i p_i) is a first-order sensitivity model stated without derivation or error bars on linearization accuracy; contributor percentages may not be reliable.
sourcesclaude-sonnet-4-20250514
Completeness 4/5
This paper presents a complete experimental demonstration of below-threshold quantum error correction. The work systematically establishes all necessary components: proper variable definitions, clear experimental protocols, comprehensive data analysis with uncertainty quantification, and validation across multiple code distances and decoding schemes. The authors address their stated objectives of demonstrating exponential error suppression (Λ = 2.14±0.02), achieving break-even performance, and implementing real-time decoding. While some technical details are relegated to supplementary information and the physical origin of certain error mechanisms remains unexplored, these represent areas for future work rather than gaps in the current claims. The core demonstration of below-threshold operation is complete and well-supported by the presented data and analysis.
+ Comprehensive experimental methodology with clear variable definitions and statistical analysis+ Systematic comparison across multiple code distances and decoding schemes with proper uncertainty propagation+ Thorough investigation of error mechanisms including detailed error budget analysis and sensitivity studies
- The physical origin of correlated error bursts limiting high-distance performance remains unexplained- Some technical details of simulation methods and decoder implementations are primarily in supplementary materials
sourcesgpt-5.4-2026-03-05
Completeness 4/5
This is a strong and largely complete experimental paper. Its main claims are not left as unsupported assertions: below-threshold behavior is evidenced by measured logical error suppression with increasing surface-code distance, beyond-break-even performance is quantified via an explicit lifetime metric, real-time decoding is evaluated using both accuracy and latency, and repetition-code experiments are used to identify a low-rate correlated-error floor. The work is careful to distinguish measured behavior from interpretation, especially where proxy metrics or simulation-based error budgets are involved.
The paper is not perfectly self-contained because a meaningful fraction of the technical support sits in the supplement, particularly for decoder training, uncertainty propagation, and simulation methodology. But those details do exist and appear well aligned with the main text. The largest remaining support gap is not in the main claim of below-threshold operation, but in mechanistic understanding of the rare correlated events that limit the highest-distance repetition-code performance. Overall, the submission is complete enough to substantiate its stated goals, with clear evidence trails and appropriately disclosed limitations.
+ Core claims are directly supported by reported measurements across multiple code distances, with quantitative uncertainties and clear operational definitions of Λ, ε_d, p_L, and latency.+ The supplementary material closes many potential completeness gaps by detailing fitting methods, uncertainty propagation, decoder configurations, simulation assumptions, circuit structure, and released datasets.+ The paper explicitly discusses limitations and edge conditions, including decoder dependence, leakage effects, drift, finite-size behavior, and the observed rare correlated-error floor.
- Several essential methodological details are deferred to the supplement, making the main text alone less self-sufficient than ideal for reproducing the key analysis pipeline.- The causal interpretation of the rare correlated bursts remains unresolved; the paper documents the events well but cannot yet support mechanistic claims about their origin.- Detection probability is used as a proxy for physical error in several places, but the paper also notes conditions where this proxy breaks down; this nuance is important and could be emphasized more consistently in the main text.- Large-scale resource and performance projections in the outlook rely on simulations and extrapolation rather than direct experimental verification.- Comparisons among decoders show performance dependence on decoder choice and prior calibration; while acknowledged, this means some headline performance metrics are not decoder-agnostic.
sourcesgemini-2.5-flash
Completeness 5/5
This paper provides a highly comprehensive experimental demonstration of quantum error correction below the surface code threshold. It rigorously presents the realization of distance-5 and distance-7 surface codes, quantifying their logical error suppression factors and demonstrating beyond break-even performance. The work effectively integrates real-time decoding, addressing a crucial requirement for fault-tolerant quantum computing, and details the trade-offs between speed and accuracy.
Furthermore, the paper thoroughly probes the limits of error correction using high-distance repetition codes, identifying and characterizing a new class of rare correlated error events that currently impose a logical error floor. The authors are transparent about the remaining challenges, such as the unknown origins of these error bursts and the substantial resource requirements for scaling, which contributes to the completeness of the scientific narrative.
+ Clear and quantitative articulation of experimental results against stated goals, providing specific performance metrics.+ Explicit discussion of limitations, challenges, and identified error mechanisms (e.g., correlated error events, error floors), demonstrating a thorough understanding of the experimental system's boundaries.+ Comprehensive support via detailed methodology, error budgeting, and extensive supplementary information on technical aspects (decoders, simulation details, uncertainty analysis).
- The paper identifies a new class of large, correlated error bursts that set a 10^-10 logical error floor, but their origins are not yet understood, posing a fundamental unknown for future scaling efforts.- While real-time decoding is demonstrated, there is a recognized 'modest reduction in accuracy' compared to offline decoders, with further work needed to integrate high-accuracy offline techniques into real-time systems.
sciencegpt-5.4-2026-03-05
Clarity 4/5Novelty 4/5Falsifiability 5/5
This is a strong and scientifically valuable experimental paper. Its main merit is not speculative theorizing but a precise empirical demonstration: logical error rates decrease with increasing surface-code distance in the reported regime, with quantitative suppression factors above unity and a distance-7 memory surpassing the best constituent physical qubit lifetime. Those are clear, testable statements backed by measurements rather than rhetorical claims. The inclusion of real-time decoding while retaining below-threshold behavior materially strengthens the paper because it addresses an operational bottleneck rather than only an offline post-processing benchmark.
In originality, the work is best understood as a substantial systems advance built from known ingredients rather than a new theoretical framework. That still counts as real novelty, because the achieved regime and integration of capabilities appear to move beyond prior demonstrations. Communication is overall strong, though some core technical details sit in the supplement and the multiplicity of decoders/metrics makes the paper dense. Overall, the submission scores highly on scientific merit: it is quantitatively falsifiable, meaningfully advances the field, and is communicated with good discipline and relatively little overclaim.
+ Central claims are tied to directly measured, quantitative observables with clear falsification routes.+ Strong novelty in integrated experimental achievement: below-threshold scaling, beyond-break-even lifetime, and real-time decoding in one platform.+ Communication is well structured, with explicit separation between measured results, simulation-supported interpretation, and forward-looking extrapolation.
- The paper relies heavily on decoder choice and decoder training details; although this is discussed, the extent to which the headline metrics depend on decoder architecture may matter for broader interpretability.- Some important methodological details are deferred to supplementary material, especially uncertainty analysis, simulation assumptions, and decoder fine-tuning procedures.- The key phrase 'below threshold' is empirically supported here through finite-distance scaling, but the paper could state even more explicitly the exact operational criterion being used, especially given known finite-size crossover issues discussed in the supplement.- The outlook extrapolations to algorithmically relevant scales are sensible but remain model-dependent and should not be read as demonstrated scalability in practice.
scienceclaude-opus-4-7
Clarity 5/5Novelty 5/5Falsifiability 5/5
This paper presents a landmark experimental result in quantum error correction: the first unambiguous demonstration of exponential logical error suppression across multiple code distance increments on a superconducting platform, with Λ = 2.14 sustained from d=3 through d=7, and a logical memory exceeding physical qubit lifetime by 2.4x. The work is rigorously documented, with extensive supplementary material covering decoder architectures, error budgets, simulation methodology, and uncertainty propagation. Open data release supports independent verification. The integration of real-time decoding and the discovery of a new ~hourly correlated error mechanism at the 10^-10 floor are both significant contributions in their own right.
From a scientific communication standpoint, the paper is exemplary: claims are precisely quantified with uncertainties, decoder choices are justified and compared, and limitations (deviation from ideal scaling, unexplained error bursts, error-budget gaps) are explicitly acknowledged rather than hidden. The work scores at the top of all three dimensions evaluated here. Remaining open questions — origin of the rare error bursts, the 20% gap between simulation and experiment, and whether Λ holds at larger distances — are clearly identified as the next experimental targets, demonstrating strong scientific self-awareness.
+ Clear, decisive empirical demonstration of below-threshold surface code operation with reproducible scaling (Λ = 2.14) confirmed across multiple decoders and replicated on two different processors.+ Identification of a new class of rare (~1/hour) correlated error bursts setting a 10^-10 floor — a concrete, actionable target for the field that would have remained hidden without distance-29 repetition codes.+ Integration of real-time decoding meeting strict throughput constraints (63μs latency, 1.1μs cycle) with only modest accuracy loss vs offline neural-network decoder, addressing a previously theoretical concern about classical co-processor scaling.
- Observed scaling exponent in error injection is ~80% of the ideal (d+1)/2, attributed to excess correlations; this deviation is acknowledged but its origin is not fully characterized, leaving uncertainty about whether further scaling will preserve Λ > 2.- The origin of the ~hourly correlated error bursts (distinct from cosmic-ray events) is explicitly stated as unknown; this is an honest acknowledgment but represents a scientific gap rather than a flaw in the paper.- The error budget overpredicts Λ by 20%, indicating the noise model does not fully capture device behavior; the unmodeled contributions are not identified.
