We present a structural, non-parametric test of quantum-error-correction syndrome streams against entire classes of noise models at once, together with a cautionary correction to a natural first version of the test. The test has two stages. Stage one is the sign stage. If a stationary binary stream is generated by a finite hidden-Markov model in Moore form, meaning state-conditional emissions, whose underlying chain is reversible in the detailed-balance sense, then the run statistics m_k = p(1^{k+1}) form a Hamburger moment sequence. Consequently the offset Hankel matrices [m_{i+j}] are positive semidefinite at every order. We prove this by detailed-balance symmetrization. Two scope corrections are essential. First, the n = 0 entry p(1^0) = 1 must be excluded: including it in the naive Hankel falsely rejects legitimate periodic models, as a deterministic two-state alternator shows. Second, the Moore qualifier is essential: a Mealy, or edge-emission, model on a reversible chain is not covered, because the induced Moore model lives on the edge chain, and the edge chain of a reversible chain is not generally reversible. A statistically signed offset Hankel therefore certifies that no Moore-form hidden-Markov model with reversible chain, of any dimension, reproduces the stream. This is a representation-class statement, narrower than an earlier draft's claim and stated exactly. Stage two is the pole stage. Prony analysis of m_k separates the two ways a sign can arise: real or rational pole structure, compatible with non-reversible finite models, versus a complex pair at irrational phase on the dominant circle, which by the theory of nonnegatively realizable sequences excludes every finite-state model. The dominance qualifier is essential, since subdominant complex pairs are unconstrained, and the protocol tests it explicitly. The pole stage requires order selection, namely numerical Hankel rank before Prony; fixed-order fitting provably hallucinates poles, as shown on an exactly solvable model. Applied to Google's 2023 surface-code memory experiment, using public data from 64 stabilizer streams with 50,000 shots each, 64/64 streams nominally sign: decisively for the typical stream, with minimum eigenvalue around -10^{-3}, tight bootstrap intervals at B = 2000, survival under a conservative Bonferroni statement, and signal an order of magnitude above the derived within-shot drift systematic; provisionally for the weakest streams, where the residual concern is bootstrap boundary behavior rather than drift. The order-selected pole stage finds purely real primary poles in 64/64 streams, and the rank scan finds no stream with a resample-stable, dominant complex pair. Thus the clock signature is absent, while an oscillatory control exhibits it at stability and dominance 1.00. The verdict is anchored by an exact model lemma: the standard detection-event mechanism, in which measurement errors fire the same detector in consecutive rounds, is a four-state chain of letter-transfer rank two with exact poles {+0.2836, -0.2336}. It signs the offset Hankel with purely real poles while being distributionally time-reversal invariant and Mealy-realizable on a reversible chain. Therefore the sign certifies the absence of a reversible Moore-form hidden realization, not a temporal arrow in the data and not the absence of all reversible-chain models. The device's syndrome statistics show exactly the pair-mechanism structure documented by the source publication by other means, and no rotation, or "clock", signature that would obstruct finite-state modeling up to the dimension bound finite data can certify. Controls demonstrate instrument power: an i.i.d. stream passes, while a hidden irrational rotation at matched rate is decisively rejected.
We present a structural, non-parametric test of quantum-error-correction syndrome streams against entire classes of noise models at once, together with a cautionary correction to a natural first version of the test. The test has two stages. Stage one is the sign stage. If a stationary binary stream is generated by a finite hidden-Markov model in Moore form, meaning state-conditional emissions, whose underlying chain is reversible in the detailed-balance sense, then the run statistics m_k = p(1^{k+1}) form a Hamburger moment sequence. Consequently the offset Hankel matrices [m_{i+j}] are positive semidefinite at every order. We prove this by detailed-balance symmetrization. Two scope corrections are essential. First, the n = 0 entry p(1^0) = 1 must be excluded: including it in the naive Hankel falsely rejects legitimate periodic models, as a deterministic two-state alternator shows. Second, the Moore qualifier is essential: a Mealy, or edge-emission, model on a reversible chain is not covered, because the induced Moore model lives on the edge chain, and the edge chain of a reversible chain is not generally reversible. A statistically signed offset Hankel therefore certifies that no Moore-form hidden-Markov model with reversible chain, of any dimension, reproduces the stream. This is a representation-class statement, narrower than an earlier draft's claim and stated exactly. Stage two is the pole stage. Prony analysis of m_k separates the two ways a sign can arise: real or rational pole structure, compatible with non-reversible finite models, versus a complex pair at irrational phase on the dominant circle, which by the theory of nonnegatively realizable sequences excludes every finite-state model. The dominance qualifier is essential, since subdominant complex pairs are unconstrained, and the protocol tests it explicitly. The pole stage requires order selection, namely numerical Hankel rank before Prony; fixed-order fitting provably hallucinates poles, as shown on an exactly solvable model. Applied to Google's 2023 surface-code memory experiment, using public data from 64 stabilizer streams with 50,000 shots each, 64/64 streams nominally sign: decisively for the typical stream, with minimum eigenvalue around -10^{-3}, tight bootstrap intervals at B = 2000, survival under a conservative Bonferroni statement, and signal an order of magnitude above the derived within-shot drift systematic; provisionally for the weakest streams, where the residual concern is bootstrap boundary behavior rather than drift. The order-selected pole stage finds purely real primary poles in 64/64 streams, and the rank scan finds no stream with a resample-stable, dominant complex pair. Thus the clock signature is absent, while an oscillatory control exhibits it at stability and dominance 1.00. The verdict is anchored by an exact model lemma: the standard detection-event mechanism, in which measurement errors fire the same detector in consecutive rounds, is a four-state chain of letter-transfer rank two with exact poles {+0.2836, -0.2336}. It signs the offset Hankel with purely real poles while being distributionally time-reversal invariant and Mealy-realizable on a reversible chain. Therefore the sign certifies the absence of a reversible Moore-form hidden realization, not a temporal arrow in the data and not the absence of all reversible-chain models. The device's syndrome statistics show exactly the pair-mechanism structure documented by the source publication by other means, and no rotation, or "clock", signature that would obstruct finite-state modeling up to the dimension bound finite data can certify. Controls demonstrate instrument power: an i.i.d. stream passes, while a hidden irrational rotation at matched rate is decisively rejected.
Abstract Quantum error-correction decoders must convert noisy syndrome records into logical decisions while operating under strict latency and reliability constraints. Minimum-weight perfect matching (MWPM) remains a strong baseline for surface-code decoding because it is fast, interpretable, and exact for its graphlike optimization problem [1–8]. However, the graphlike independent-error approximation can become mismatched when syndrome records contain temporal drift, measurement corruption, crosstalk, bursts, or mixtures of mechanisms [9–18]. We introduce SAGE-QEC, a failure-aware selective decoder that retains nominal PyMatching as its default decision and learns when an observable syndrome structure warrants an alternative candidate. SAGE-QEC uses train-frozen structure thresholds, a candidate ranker, validation-only policy selection, and a paired blind evaluation. Across 45 simulator conditions, SAGE-QEC reduced the aggregate logical error rate from 27.1285% for PyMatching to 24.8320%, corresponding to an 8.47% relative reduction. The campaign used rotated surface-code memory circuits at distances 3, 5, and 7, nominal circuit parameters 0.002, 0.003, and 0.005, five seeds, and four test regimes withheld from training: Burst-3, Temporal Drift, Crosstalk, and Mixed Noise. The test set contained 81,000 paired shots. All twelve distance-by-regime aggregate comparisons favored SAGE-QEC in the executed campaign, with paired bootstrap intervals below zero for SAGE-QEC minus PyMatching. The result is evidence of a reproducible simulator-level improvement, not a claim of universal decoder superiority. SAGE-QEC currently uses logical-output proxy candidates rather than a complete physical correction search, and its intervention rate remains high under distribution shift. It has not yet been compared with a fully calibrated correlated-noise space-time PyMatching baseline or hardware calibration data. We therefore characterize SAGE-QEC as a strong qualification-stage prototype and release a complete audit protocol for the next industrial validation stage.
Whether financial data require a quantum-probabilistic description has been debated since Khrennikov proposed Bell-type and total-probability tests for markets, yet the field still lacks an instrument whose verdicts can be trusted in either direction. We first show that Bell-type and Leggett–Garg inequalities computed on passively recorded price data cannot establish non-classicality by violating a classical bound: by Fine's theorem, jointly recorded variables always admit a classical joint distribution, so apparent violations arise only from combining different conditioning contexts, which can test specific classical models but not the need for a quantum description. We then construct an alternative instrument within this boundary: a family of quantum-like models (a two-dimensional collapse chain and general hidden quantum Markov models of dimension two and three) is compared out of sample with a deliberately strong classical class (Moore and Mealy hidden Markov models with up to six states, including the classical limits of the quantum models, and an observation-driven GARCH-type model). A quantum member counts as evidence only if it beats every classical candidate. Before reading real data, we calibrate the instrument in both directions on synthetic data with known ground truth. On one year of 15-minute data it detects quantum structures matched to its members down to about 0.001 log-likelihood units per bar, while producing no false detections in 20 classical simulations read under the same rule, nor in 60 earlier ones under a less strict rule; the real-data conclusions survive a larger classical class and a doubled optimisation budget. Calibration also exposes five ways in which quantum advantage can be manufactured from classical data. Applied to seasonally adjusted gold and silver data at two time scales, the instrument detects no quantum-like advantage; every 95% upper bound is negative. A pre-registered out-of-sample test on 2026 gold data was likewise not supported. The contribution is a tested ruler with stated resolution, not a claim about the ontology of markets. Code, frozen measurement protocol and all results: https://doi.org/10.5281/zenodo.23449779 (version 1.1). Pre-registration: https://doi.org/10.5281/zenodo.23400902; results record: https://doi.org/10.5281/zenodo.23401850. This paper revises the author's earlier claim that the interference statistic κ indicates non-classicality in gold (https://doi.org/10.5281/zenodo.19456823).
Exact physical memory is an unstable target: a process of Hankel rank one can be arbitrarily close to processes requiring larger classical and quantum carriers. We make this obstruction quantitative and then identify the extra structure under which a physical finite-sample upper theorem is possible. First, on a stationary binary Markov family whose mixing gap is uniformly at least one half, every estimator of exact linear rank, minimum classical memory, or minimum quantum Hilbert dimension has worst-case error at least 3/8 at local separation Θ([N(L−1)]−1/2) from the rank-one model. The smallest nonzero singular value of an explicit Hankel core is proportional to the same parameter. Thus a spectral or dynamical margin is necessary even under rapid mixing. Second, finite Hankel data determine a minimum linear realization only up to similarity, whereas nonnegativity and complete positivity are not invariant under a generic similarity. We give an exact two-state counterexample and then formulate the missing assumption as a calibrated physical core. For an invertible core H, all shifts are recovered in one basis by A~a=H~−1H~a. An explicit perturbation theorem, followed by Choi positive-part projection and one common trace-preserving correction, returns a genuine quantum instrument. A direct-query finite-sample corollary controls its error uniformly over adaptive policies and exposes every dependence on core conditioning, physical calibration, horizon, and confidence. The resulting linear-in-horizon coefficient is first-order sharp: a one-state classical process attains the exact discrepancy 1−(1−Δ/2)L at one-step diamond radius Δ. Third, we prove a compact-state classicalization theorem. Any finite-dimensional physical memory described by a compact base-norm state space has, at horizon L and operational error η>0, a classical realization whose number of states is the seeded covering number at radius 2η/L, uniformly over adaptive policies. For a d-dimensional quantum instrument this gives, for d≥2, (1+η2(1−d−1)L)d2−1 states, with no extra transient state; one state suffices for d=1. The ambient exponent d2−1 is optimal for universal trace-norm discretization, while a compact branch-invariant reachable set replaces it by its own metric entropy. Consequently, an unbounded exact quantum/classical memory separation cannot remain unbounded at fixed d, L, η. Robust quantum advantage must instead be quantified as a joint precision–horizon tradeoff. Deterministic verification code audits the mixing lower bound, the similarity-gauge counterexample, common-shift recovery, physical repair, and finite-net simulation.
This preprint studies finite-shot reconstruction of hidden quantum factor structure from global calibration-free sequence statistics without assuming an internal tensor-product cut in advance. A probability-only complex-projective 3-design self-test is used as an imported anchor for the active Hilbert/Jordan geometry. From this anchor, global control maps are reconstructed with explicit confidence bounds, including a weighted estimator for unequal shot counts and missing reference settings whenever the surviving frame remains full rank. The main certification result is a whole-commutant-energy lower bound. For reconstructed controls, the sum of the lowest commutator singular energies yields a basis-free lower bound on the aggregate projective Hilbert–Schmidt distance to every common hidden matrix factor of a specified multiplicity. Combined with exact Choi-to-projective-unitary conversion and finite-shot confidence balls, this produces a one-sided lower bound on distance to the complete common-factor model class. An explicit exact factor witness gives the complementary PROXIMAL-RECOVER certificate; otherwise the procedure returns ABSTAIN. The paper also proves that exact generated-algebra / tensor-product-structure type is not uniformly identifiable from finite samples without a structural separation margin. For the fixed two-qubit 60-state stabilizer benchmark, the paper provides an anchor-inclusive raw-count validation and an exact-rational / outward-rounded one-sided trust path from integer counts through anchor and control confidence radii to factor RECOVER/REJECT. The scope is deliberately operational: the result certifies active factor structure supported by the augmented data and does not claim identification of arbitrary inactive spectator degrees of freedom or a uniquely true microscopic decomposition.
An exact operational selector specifies which candidate maps are nonnegative on declared preparation and effect resources. This paper studies finite-count calibration of that resource-dependent question for a qubit apparatus with an inert qubit reference, an arbitrary complex physical preparation, and an arbitrary physical binary detector, relative to a trusted common calibration frame. Two preparation-stabilizer settings and six separable detector probes control the statistic of the affine family (1 − θ)id + θT, 0 ≤ θ ≤ 1, uniformly over all compatible state–effect pairs. The explicit radius is 3h + √(2s), where h is a simultaneous detector calibration width and s is a one-sided bound on Bell infidelity. It yields simultaneous inner and outer confidence sets for a one-parameter slice of the declared safe-map class. A fixed allocation of 200,000 trials per calibration setting, totaling 1,600,000 calibration trials, gives false-implication probability at most 0.005 and transpose-exclusion power at least 0.996 at the specified physical alternative. The calibration confidence budgets are 0.001 for preparation and 0.004 for the detector. This exclusion calculation uses no target stream; the acquired calibration circuits can use identity wiring. Coverage and prospective power remain separate statements. The acquisition contract uses fixed-length iid Bernoulli streams with their stated physical means and permits coupling between whole streams. The numerical example sets external-reference systematic errors to zero, and its trial budget excludes establishing reference trust and accuracy. If the preparation and detector references are not linked, the paper constructs two physical identity-wired apparatus worlds with identical complete record laws under the specified independent acquisition construction but different safety conclusions on the fixed target cone pair. A secondary nearest-center theorem supplies unconditional error control for the joint event of a wrong and nonabstaining candidate label under an additional finite-context target-law contract. Its nonidentity candidates are not completely positive maps, and their ideal probabilities lie below the nominal detector’s ordinary positive-output floor. That target law is an explicit model assumption; the label theorem is not a conditional-correctness or posterior statement. The results are analytic and prospective. The number of settings and trial allocations are sufficient constructions, with no empirical success-rate, minimality, optimal sample-complexity, or general certified count-to-feasibility implementation claim. The scope is one declared qubit apparatus, target preparation ray and binary effect cone, and affine map slice. The common complex operator description, trusted reference relationship and physical calibration probability law are assumptions. Exact finite checks and an assessed supporting Lean subset accompany the manuscript. That subset covers 326 audited named declarations: 240 theorems and 86 definition/type interfaces. Analytic binomial coverage, concentration and acquisition bridges, independent-product-law construction and normalization, and the complete noisy-power chain remain outside the accepted Lean subset. That subset does not establish a certified numerical feasibility evaluator, a complete physical two-sector implementation, or an ordinary-channel realization. AI assistance and the dated review/correction provenance are disclosed in the manuscript and companion records. No experimental dataset, external human peer review, or end-to-end formal verification is claimed.
Abstract We study decoder-representation sensitivity in simulated rotated surface-code Z-memory circuits with local spatially correlated XX pair events. At distance 3, ordinary minimum-weight perfect matching (MWPM) produced 4,610 versus 1,260 logical failures at p_corr = 10^-4 and 5,437 versus 2,239 at p_corr = 2 × 10^-4 on the same 500,000-shot samples using original-order versus canonical-sorted detector error models (DEMs). Across 50 additional seeded permutations per condition, failure counts varied widely. Canonical sorting tied for the minimum in this finite tested set; it is not claimed to be universally optimal. The ordering experiment reorders complete error instructions in a flattened DEM, preserving their probabilities and absolute detector and logical targets. PyMatching documents retaining the first-added logical-label set when parallel graph edges are merged; we therefore interpret the findings as a workflow-specific reproducibility result, not a newly discovered general software defect. At distances 5 and 7, correlation-aware MWPM had lower observed failure rates in all eight tested conditions, with six paired contrasts remaining significant after Bonferroni correction. An exact maximum-a-posteriori calculation gave a 0.2367% Bayes risk for the flattened-DEM distribution at p_corr = 10^-4; this is optimal for that modeled distribution only. This bounded classical simulation study proposes no new decoder, estimates no threshold, and provides no hardware evidence. Some historical source-to-result provenance remains incomplete, as disclosed in the manuscript. Code and evidence: https://github.com/tgiwn/QEC-surface-code-correlated-noise/tree/106d32e815c5f6009941afc36d3f65be7ecf35a3 Status: This is a preprint and has not undergone peer review.