Pith. sign in

REVIEW 3 major objections 6 minor 17 references

A wrong cross-boundary guess in speculative window decoding is contained to the single window it poisons; the propagated effect decays exponentially with commit width, so verified speculation adds no error floor.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 08:12 UTC pith:BJL7TZML

load-bearing objection The verifier side is a real contribution, but the formal blast-radius bound is conditional on a reduction the paper itself refutes — the 'no error floor' conclusion is currently a conjecture. the 3 major comments →

arxiv 2607.13062 v2 pith:BJL7TZML submitted 2026-07-07 quant-ph cs.AR

The verifier side of speculative window decoding: a predictability bracket, a machine-checked blast-radius bound, and a decoder-agnostic recover loop

Rylan Malarchick This is my paper
classification quant-ph cs.AR PACS 03.67.Pp
keywords speculative window decodingquantum error correctionsurface codedecoder latencyminimum-weight perfect matchingblast radiusrestart policydecoder-agnostic wrapper
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper builds the verifier half of speculative window decoding, the part that checks whether a guessed cross-window decoding decision was right and what to do when it was not. It brackets the best achievable predictor accuracy at about 0.999 with a three-round receptive field, leaving only small headroom over existing predictors, and it bounds the worst-case temporal blast radius: a wrong guess decays exponentially with the commit width—the distance in rounds from the incoming boundary to the outgoing cut—so at the standard width it is two orders of magnitude below the logical error rate. The probability core of that bound is machine-checked, but the modeling reduction it relies on is then tested shot by shot and falsified; the real propagation mechanism is a global minimum-weight re-pairing, not a faulty-path percolation. A compiler pass derives the restart policy from the measured numbers, and a runtime executor shows the predict-verify-recover loop recovers exactly and removes the serial commit-chain stall up to a penalty of order 10⁻⁵. Re-running the structural findings with a second, algorithmically distinct decoder shows the wrapper and the locality/containment phenomenology are decoder-agnostic.

Core claim

A mispredicted cross-boundary decision is contained to the single window it enters: the probability it propagates to a later window decays exponentially in the commit width W, satisfying Pprop(W) ≤ C d² λ^W with measured λ≈0.16–0.23 (rigorous rate below 1). At the standard width W=d, propagation is 1.5×10⁻⁷ at d=7, two orders below the logical error rate, so verified speculation adds no error floor. The boundary decision is local—three rounds of syndrome on each side reach about 0.999 accuracy. The paper then falsifies the faulty-path modeling assumption behind the bound and shows the real mechanism is a global minimum-weight re-pairing. A compiler pass turns the measured numbers into a rest

What carries the argument

The load-bearing object is the temporal blast-radius operator: perturb the incoming (sink) boundary of one decoding window, re-decode with minimum-weight matching, and read whether the outgoing boundary at distance W changes. The probability core is a machine-checked union-bound tail—the probability that some faulty boundary-to-cut path exists is at most the path count times q^W—which supplies the exponential decay. On the predictor side, a radius-R local matching decoder that treats the window's open boundary exactly as a window decoder does yields the achievable-accuracy bracket. A compiler pass over an abstract window dependency graph, using cost = predictability × slack − blast radius, c

Load-bearing premise

The load-bearing premise is that a misprediction can only corrupt the next window when faulty error events form an unbroken chain from the bad boundary all the way to the far cut. The paper tests exactly that premise shot by shot and finds it false for the actual propagating shots, so the formal guarantee currently rests on a hypothesis the paper itself refutes as a mechanism.

What would settle it

Re-run the single-bit-flip operator at code distance 7, commit width 7, noise 10⁻³, with enough shots to distinguish Pprop ≈ 1.5×10⁻⁷ from the logical error rate ≈ 1.5×10⁻⁵; if propagation reaches or exceeds the logical error rate, the no-error-floor claim fails. A single propagating shot whose changed matched edges contain no re-route reaching the far cut would also contradict the claimed mechanism.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Speculative window decoding can hide decoder latency without degrading the logical error rate: a wrong boundary guess is contained to the poisoned window, so the speedup is not paid for in reliability.
  • The accuracy lever for future predictors is receptive field, not predictor cleverness: three rounds of syndrome on each side of the cut nearly saturates achievable accuracy, so better local predictors buy only a few diffuse points.
  • The restart policy is derivable from measurements: restart the poisoned window plus its corner-adjacent neighbor, a choice that is robust because the temporal and spatial couplings differ by more than six orders of magnitude.
  • An N-window chain runs at the architecture's critical-path limit—16.0 out of 16 for the tested chain—up to a per-boundary restart penalty of about 10⁻⁵, even when predictability is degraded to 0.7.
  • The predict-verify-recover wrapper is decoder-agnostic: any decoder plugs in behind the verify step, and the structural results (locality, containment, radius one) transfer, while the absolute magnitudes and the min-weight mechanism do not.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The stated next step—proving that propagation needs Ω(W) matching-weight subsidy to beat the local absorb—would explain the measured exponential decay without relying on the falsified faulty-path reduction; until then, the formal bound's hypothesis is a known gap.
  • The near-tie degeneracy at high noise (about 40% of decisions flip under a small weight perturbation at p=10⁻²) suggests the boundary readout is inherently unstable near threshold, which may cap how predictable any single boundary can be at realistic noise.
  • The speedup claims are critical-path latency, not wall-clock time; the real benefit on hardware depends on the fraction of total runtime that is commit-chain latency, so a bare-metal deployment is the missing measurement.
  • The predictability bracket is internal to the reconstructed harness; a general ceiling over all predictors and all decoders remains open, so the headroom numbers should be read as conservative for this task, not as a fundamental limit.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper builds the verifier side of speculative window decoding on a reconstructed SWIPER harness (Stim surface code, MWPM). It reports three main threads: (1) a predictability bracket showing the cross-boundary decision is local and SWIPER is near the achievable accuracy; (2) a worst-case temporal blast-radius bound Pprop(W) ≤ C d² λ^W, whose probability core is machine-checked in Lean 4 but conditional on a faulty-path reduction, which the paper then tests and falsifies shot-by-shot; (3) a compiler pass and a runtime executor that derive and exercise SWIPER's restart policy, plus a union-find cross-check for decoder-agnostic claims. The paper is unusually explicit about its limitations: Sec. 4.4 refutes the very reduction that the Lean theorem consumes, and Sec. 8 states the matching-weight replacement bound is not proved. However, the abstract and conclusion still advertise the blast-radius bound and the 'no error floor' conclusion as established, which is stronger than what the evidence and the formal development actually support.

Significance. If the containment claim were established, this would be a useful step for speculative window decoding: it would show that a wrong boundary guess is temporally contained, that the restart policy is derivable from measurable quantities, and that the predict-verify-recover wrapper is decoder-agnostic. The paper has real strengths: the Lean development is machine-checked with no sorry, the harness is reproducible from fixed seeds, the calibration gate against SWIPER's reported accuracy is a good practice, the second-decoder check is a serious robustness test, and the authors test and falsify their own modeling hypothesis rather than hiding it. These methodological strengths are significant. The central limitation is that the headline 'no error floor' conclusion is not currently proved for the actual mechanism; it rests on a conditional theorem whose hypothesis is empirically false and on a fitted, extrapolated exponential with no error analysis.

major comments (3)
  1. [Sec. 4.1, 4.2, 4.4; Table 1] The advertised result—'the temporal blast radius is one, and verified speculation adds no error floor'—is not established by the formal part. The Lean theorem in Sec. 4.2 is conditional on the reduction that poison reaching the cut implies a faulty boundary-to-cut path of length at least W. Sec. 4.4 tests exactly this predicate; Table 1 gives P(reach_all | propagation) = 0.071 over the full envelope and 0.005 at W ≥ 2. Thus the hypothesis of the machine-checked containment theorem is false for the actual decoder, and the theorem cannot be instantiated. The paper acknowledges this in Sec. 8 and defers a matching-weight bound to future work, so the central 'no error floor' claim currently rests on an unproved conjecture plus a fitted curve, not on the Lean proof. The abstract and conclusion should be revised to state this conditional/unproved status, or a proof of the matching-weight bound
  2. [Eq. (1), Fig. 3, Sec. 4.2] Eq. (1) is presented as the containment bound, with λ quoted as measured (0.16, 0.19, 0.23 at d = 7, 9, 11). Sec. 4.2 then says the reduction is 'backed by the empirical certificate of Eq. (1)'. This is circular: the same decay data are used both to fit the exponential and to support the hypothesis from which the exponential bound is supposed to follow. No error bars, goodness-of-fit statistics, or holdout validation are reported for the fits in Fig. 3. The headline values at W = d (e.g., 1.5×10^-7 at d = 7) are extrapolations beyond the largest measured W in Fig. 3 for the larger distances, with no uncertainty. Since the architectural conclusion is exactly the comparison with the logical error rate at W = d, this missing error analysis is load-bearing.
  3. [Table 2, Sec. 4.4] The low-noise support for the claimed exponential containment is thin. At p = 10^-3, the regime used for the W = d claims, only n = 24 propagating events are reported, and the text calls the row 'indicative'. The statement in Sec. 4.4 that 'the exponential containment is real' is therefore not supported by strong data at the operating point. Moreover, the proposed matching-weight mechanism ('propagation needs Ω(W) subsidy to beat the local absorb') is explicitly marked as not proved in Sec. 8. The paper should either provide confidence intervals, more shots at p = 10^-3, and a direct test of the matching-weight subsidy, or explicitly downgrade this sentence to a conjecture consistent with limited data.
minor comments (6)
  1. [Eq. (1)] The symbols C and c in Eq. (1) are not defined; 'deg2 cp' is hard to parse. Please define every symbol and specify the domain of validity of the inequality.
  2. [Fig. 3] The stars at W = d lie outside the measured W range for the larger distances. Please show the extrapolation with uncertainty bands or separate the measured points from the extrapolated ones.
  3. [Sec. 6] The term 'head-C blast-radius term' is used without prior introduction; the reader is left to infer the relation to 'head-A' and 'head-B' in Sec. 3. Please define the terminology consistently.
  4. [Table 2] Report confidence intervals for the medians and proportions, especially for the n = 24 row at p = 10^-3. The pooled numbers are also confounded because 94% of propagating events occur at p = 10^-2; a rate-stratified presentation is needed.
  5. [Sec. 4.1] The statement that 'flip_all dominates' the localized budget-K adversary is only empirically verified, not proved. Since the bound is advertised as worst-case, please state explicitly that this domination is part of the empirical rather than the formal component.
  6. [Abstract and Conclusion] The abstract and conclusion say 'we establish a worst-case temporal blast-radius bound' and 'no error floor' without the qualifications given in Sec. 8. Please align the summary statements with the actual status of the result.

Circularity Check

2 steps flagged

The 'no error floor' bound is a fit to the same Pprop curve it claims to bound; the Lean theorem's reduction hypothesis is justified by that same fit and is then refuted, so the central formal claim rests on an extrapolation plus an unproved matching-weight conjecture.

specific steps
  1. fitted input called prediction [Sec. 4.1, Eq. (1), Fig. 3]
    ""The propagation probability decays exponentially in the commit width W, Pprop(W)≤Cd2λW, λ=√deg2cp<1... The measured rate is λ=0.16, 0.19, 0.23 at d=7, 9, 11... At the standard commit width W=d the failure probability is 1.5×10−7... The temporal blast radius is one, and verified speculation adds no error floor." Fig. 3 caption: "line = fitted decay, star = extrapolated P_prop at standard W=d"."

    Eq. (1) is not an unconditional derivation from the Lean core for the real decoder: the theorem is conditional on a reduction hypothesis. The exponential rate and d² prefactor are fit to the measured Pprop(W) curve, and the W=d 'failure probability' and 'no error floor' are extrapolations of that fitted curve, not independent verifications. Thus the headline containment prediction is the fitted model evaluated at W=d, which is the same observable used to determine the fit.

  2. other [Sec. 4.2]
    ""What is not proved from the decoder, and is carried as an explicit hypothesis backed by the empirical certificate of Eq.(1), is the reduction itself: that poison reaching the cut implies a faulty boundary-to-cut path of length at least W, together with the surface-code constants.""

    The reduction hypothesis is the premise from which the exponential containment bound is supposed to follow, and Eq. (1) is that very containment bound. Using the empirical certificate of Eq. (1) as the support for the hypothesis is affirming the consequent: the conclusion is used to justify its own premise. The paper then falsifies the hypothesis in Sec. 4.4 (reach_all survives only 0.071 of propagating shots), so the Lean theorem cannot be instantiated for the actual mechanism, and Sec. 8 concedes the matching-weight replacement is not proved.

full rationale

Most of the paper is careful, self-contained empirical reconstruction with an honest, explicitly stated falsification, and there is no load-bearing self-citation or renaming issue. The circularity is concentrated in the central temporal blast-radius claim. The machine-checked Lean core is a correct conditional theorem, but the condition it depends on—the faulty boundary-to-cut path reduction—is 'backed by the empirical certificate of Eq.(1)', i.e., by the very exponential containment conclusion it is supposed to imply. Sec. 4.4 then refutes the reduction shot by shot, and Sec. 8 states that the replacement matching-weight bound 'is not proved'. Accordingly, the advertised 'blast radius is one / no error floor' result is not established by the formal proof; what remains is the fitted exponential decay of Fig. 3, extrapolated to W=d. That makes the central prediction a fit to the same observable it claims to bound, with the formal machinery conditional on a refuted premise. The paper's own limitation statements are weighed here and support, rather than mitigate, the partial-circularity finding: the conclusion reduces to an extrapolation and a conjecture, not to the Lean theorem.

Axiom & Free-Parameter Ledger

2 free parameters · 6 axioms · 0 invented entities

The central bound rests on two fitted quantities (λ, C) and on a reduction axiom the paper itself falsifies; the speedup and no-error-floor conclusions are thus partly empirical rather than derived. No new physical entities are introduced; "poison" and "re-pairing" are descriptive mechanisms.

free parameters (2)
  • λ (exponential decay rate) = 0.16, 0.19, 0.23 for d = 7, 9, 11 at p = 1e-3
    Fit to measured Pprop(W) in Fig. 3; then used to extrapolate Pprop(W = d) and to assert containment/no error floor.
  • C (prefactor in Pprop ≤ C d² λ^W) = order-one, not specified
    Absorbed with the d² prefactor to match the measured curve; no closed-form value is given.
axioms (6)
  • ad hoc to paper Poison reaching the output cut implies a faulty boundary-to-cut path of length at least W, plus surface-code constants (the "reduction").
    Sec. 4.2 makes it the explicit hypothesis of the Lean-certified bound; Sec. 4.4 falsifies it shot-by-shot (canonical predicate survival 0.071).
  • ad hoc to paper flip_all (every sink bit flipped) dominates any localized budget-K adversary cluster.
    Sec. 4.1 uses this to extend containment to the worst-case adversary; verified empirically, not proved.
  • domain assumption Stim rotated surface-code memory circuits with uniform circuit-level depolarizing noise at rate p on all channels model the physical error process.
    Sec. 2 harness; standard in QEC simulation but a modeling choice.
  • domain assumption The Leviathan et al. speculative-decoding speedup formula maps boundary accuracy to speedup with free depth γ ≤ 32.
    Sec. 3 maps headroom to speedup; the paper states the ratios are optimistic under linear-chain rollback.
  • domain assumption MWPM via PyMatching is the reference decoder and its cross-cut matched-edge parities define the dependency bits.
    Sec. 2; all boundary predict/verify claims are relative to this decoding rule.
  • standard math Peierls/union-bound tail, path-count bound, and graph-degree constants used for the Lean theorem.
    Sec. 4.2; machine-checked in Lean 4/Mathlib, treated as background.

pith-pipeline@v1.3.0-alltime-deepseek · 10395 in / 19444 out tokens · 179567 ms · 2026-08-02T08:12:12.658769+00:00 · methodology

0 comments
read the original abstract

Speculative window decoders hide quantum error-correction decoder latency by guessing the cross-boundary decisions that link adjacent decoding windows, running downstream work on the guess, and verifying lazily. SWIPER and ARTERY each build one predictor, about 90% accurate; neither built the verifier side. We build it on a reconstructed SWIPER harness (Stim rotated surface code, minimum-weight matching). A predictor-only bracket shows the cross-boundary decision is local, the achievable accuracy reaching about 0.999 within three rounds, with small, diffuse headroom over SWIPER. We establish a worst-case temporal blast-radius bound, its probability core machine-checked in Lean4 and conditional on a modeling reduction we then test: a misprediction's effect decays exponentially in the commit width, so the radius is one and speculation adds no error floor. We falsify that reduction shot by shot and find the real mechanism, clearest at near-threshold noise, is a global minimum-weight re-pairing. A compiler pass derives SWIPER's restart policy from these numbers; a runtime executor confirms on the harness that the loop recovers exactly and removes the serial commit-chain stall up to a small penalty. A second decoder (union-find) settles which results are decoder-agnostic: the predict-verify-recover wrapper and the structural phenomenology, while the absolute magnitudes and the min-weight mechanism are matching-specific.

Figures

Figures reproduced from arXiv: 2607.13062 by Rylan Malarchick.

Figure 1
Figure 1. Figure 1: Calibration gate. The reconstructed harness repro [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Worst-case temporal poison propagation Pprop(W) decays exponentially in the commit width W, with rate λ < 1 at every distance and a d 2 prefactor. At W = d the failure sits well below the logical error rate, so the blast radius is one. The finding is that the corruption is contained to the poisoned window. The propagation probability decays exponentially in the commit width W, Pprop(W) ≤ C d2 λ W , λ = q d… view at source ↗
Figure 4
Figure 4. Figure 4: Decoder-relativity under a second decoder (union [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

17 extracted references · 12 linked inside Pith

  1. [1]

    Blunt, Archibald Ruban, Jimmy McDaniel, Joseph A

    Laura Caune, Luka Skoric, Nick S. Blunt, Archibald Ruban, Jimmy McDaniel, Joseph A. Valentine, et al. Demonstrating real-time and low-latency quantum error correction with su- perconducting qubits.arXiv preprint, 2024. arXiv:2410.05202. 7

    Pith/arXiv arXiv 2024

  2. [2]

    Quantum error correction below the surface code threshold.Nature, 638:920–926, 2025

    Google Quantum AI and Collaborators. Quantum error correction below the surface code threshold.Nature, 638:920–926, 2025. arXiv:2408.13687

    Pith/arXiv arXiv 2025

  3. [3]

    Francesco Battistel, Christopher Chamberland, Kauser Johar, Ramon W. J. Overwater, Fabio Sebastiano, Luka Skoric, Yosuke Ueno, and Muhammad Usman. Real-time decoding for fault-tolerant quantum computing: Progress, challenges and outlook.Nano Futures, 7:032003,

  4. [4]

    Barbara M. Terhal. Quantum error correction for quantum memories.Reviews of Modern Physics, 87:307–346, 2015. arXiv:1302.3428

    Pith/arXiv arXiv 2015

  5. [5]

    Browne, Kenton M

    Luka Skoric, Dan E. Browne, Kenton M. Barnes, Neil I. Gillespie, and Earl T. Campbell. Parallel window decoding enables scalable fault tolerant quantum computation.Nature Communications, 14:7040, 2023

    2023

  6. [6]

    Topological quantum memory

    Eric Dennis, Alexei Kitaev, Andrew Landahl, and John Preskill. Topological quantum memory. Journal of Mathematical Physics, 43(9):4452– 4505, 2002. arXiv:quant-ph/0110143

    Pith/arXiv arXiv 2002

  7. [7]

    Chadwick, Sarang Joshi, Gokul Subramanian Ravi, Yanjing Li, and Fred- eric T

    Joshua Viszlai, Jason D. Chadwick, Sarang Joshi, Gokul Subramanian Ravi, Yanjing Li, and Fred- eric T. Chong. SWIPER: Minimizing fault- tolerant quantum program latency via specula- tive window decoding. InProceedings of the 52nd Annual International Symposium on Computer Architecture (ISCA), 2025. arXiv:2412.05115

    Pith/arXiv arXiv 2025

  8. [8]

    ARTERY: Fast quantum feedback using branch prediction

    Wuwei Tian, Liqiang Lu, Siwei Tan, Yun Liang, Tingting Li, Kaiwen Zhou, Xinghui Jia, and Jian- wei Yin. ARTERY: Fast quantum feedback using branch prediction. InProceedings of the 52nd An- nual International Symposium on Computer Ar- chitecture (ISCA), 2025

    2025

  9. [9]

    Fastinferencefromtransformersviaspecula- tive decoding

    Yaniv Leviathan, Matan Kalman, and Yossi Ma- tias. Fastinferencefromtransformersviaspecula- tive decoding. InProceedings of the 40th Interna- tional Conference on Machine Learning (ICML),

  10. [10]

    Accelerating large language model decoding with speculative sampling.arXiv preprint, 2023

    Charlie Chen, Sebastian Borgeaud, Geoffrey Irv- ing, Jean-Baptiste Lespiau, Laurent Sifre, and John Jumper. Accelerating large language model decoding with speculative sampling.arXiv preprint, 2023. arXiv:2302.01318

    Pith/arXiv arXiv 2023

  11. [11]

    Promatch: Extending the reach of real-time quantum error correction with adaptive predecoding

    Narges Alavisamani, Suhas Vittal, and Moinud- din Qureshi. Promatch: Extending the reach of real-time quantum error correction with adaptive predecoding. InProceedings of the 29th ACM In- ternational Conference on Architectural Support for Programming Languages and Operating Sys- tems (ASPLOS), 2024. arXiv:2404.03136

    Pith/arXiv arXiv 2024

  12. [12]

    Pinball: A cryogenic predecoder for surface code decoding under circuit-level noise.arXiv preprint, 2025

    Alexander Knapen, Guanchen Tao, Jacob Mack, et al. Pinball: A cryogenic predecoder for surface code decoding under circuit-level noise.arXiv preprint, 2025. arXiv:2512.09807

    Pith/arXiv arXiv 2025

  13. [13]

    Stim: a fast stabilizer circuit sim- ulator.Quantum, 5:497, 2021

    Craig Gidney. Stim: a fast stabilizer circuit sim- ulator.Quantum, 5:497, 2021. arXiv:2103.02202

    Pith/arXiv arXiv 2021

  14. [14]

    Sparse blos- som: correcting a million errors per core sec- ond with minimum-weight matching.Quantum, 9:1600, 2025

    Oscar Higgott and Craig Gidney. Sparse blos- som: correcting a million errors per core sec- ond with minimum-weight matching.Quantum, 9:1600, 2025. arXiv:2303.15933

    Pith/arXiv arXiv 2025

  15. [15]

    Fowler, Matteo Mariantoni, John M

    Austin G. Fowler, Matteo Mariantoni, John M. Martinis, and Andrew N. Cleland. Surface codes: Towards practical large-scale quantum compu- tation.Physical Review A, 86:032324, 2012. arXiv:1208.0928

    Pith/arXiv arXiv 2012

  16. [16]

    The Lean mathemati- cal library

    The Mathlib Community. The Lean mathemati- cal library. Proc. 9th ACM SIGPLAN Int. Conf. on Certified Programs and Proofs (CPP), 2020

    2020

  17. [17]

    Nickerson

    Nicolas Delfosse and Naomi H. Nickerson. Almost-linear time decoding algorithm for topological codes.Quantum, 5:595, 2021. arXiv:1709.06218. 8

    Pith/arXiv arXiv 2021