PaperlayerView exact revision on arXiv

Source-linked AI summary

Error-Rate Reduction in LDPC Decoding via Bit-Aligned Temporal Reinforcement in Parallel Probabilistic-Bit Dynamics

Naoya Onizawa, Takahiro Hanyu

arXiv:2609.12389v1cs.ITcs.ETmath.ITphysics.comp-ph

TL;DR

Highly parallel p-bit updates can change collective dynamics, raising the question of which temporal state helps LDPC inference. The paper tests additive reuse of each bit’s saturated response and reports improved decoding outcomes, with trajectory analyses linking the advantage to correct-basin acquisition and stability.

  • Problem

    Highly parallel p-bit updates can alter collective dynamics, motivating investigation of which form of bit-local temporal state improves LDPC inference.

    Source: §I · p. 1

  • Method

    The study tests additive saturated-response reinforcement against gain, normalization, shuffling, and tuned binary-state feedback controls.

    Source: §VIII · p. 9

  • Results

    Additive saturated-response memory improves representative-code BER and FER, promotes correct-basin acquisition, and improves post-acquisition stability.

    Source: §VIII · p. 9, §IX · p. 10

  • Takeaways & Limitations

    The results motivate co-designing the retained variable and its response-path reinjection around computational structure.

    Source: §IX · p. 10

    Supplemental MAX-CUT and random 2-SAT boundary controls do not support a generic temporal-state advantage under the tested conditions.

    Source: §VIII · p. 10

Abstract

from arXiv · show

Probabilistic bits (p-bits) provide a physical and algorithmic primitive for stochastic inference, but highly parallel updates can alter their collective dynamics. We study the decoding of random-regular (3, 6) low-density parity-check (LDPC) codes using Jacobi-type p-bit annealing with stochastic partial activation. An additive response-path rule stores each bit's saturated response and reuses it before stochastic readout. High-statistics simulations with independent parameter optimization for each method show pooled bit-error-rate reductions of 33.5%, 74.8%, and 81.8% relative to memoryless probabilistic simulated annealing (pSA) for representative codes of block lengths 96, 192, and 288, respectively. Static gain, normalized averaging, response shuffling, and same-bit binary-state feedback with only its coefficient tuned under the same nonmemory parameters do not reproduce the full saturated-response benefit. Trajectory analysis links the improvement to acquisition of the channel-consistent valid-codeword basin and enhanced post-acquisition stability; the acquisition advantage persists from random and channel-hard-decision starts under the tested conditions. Across all 30 independent code realizations, fixed additive parameter-and-readout packages achieve lower bit- and frame-error rates than separately optimized pSA-specific packages, with neither package retuned for individual codes. These results show that the computational effect of temporal state depends on the retained quantity and its reinjection into stochastic dynamics.

I. INTRODUCTION

The paper asks which bit-local temporal state can improve LDPC inference under parallel p-bit dynamics. It tests additive reuse of each bit’s saturated response against alternative memory mechanisms and situates that question within parallel-update and LDPC-decoding research.

I. INTRODUCTION

  • Parallel schedules and simultaneous commitment can alter convergence even when the instantaneous p-bit response law is unchanged.Prior temporal methods retain different variables and reinject them in different ways, so a single memory coefficient does not summarize their transition kernels.Source: §I · p. 1
  • The work studies Jacobi-type field evaluation with stochastic partial activation, unlike a prior pipelined FPGA decoder that preserves sequential logical update order.Belief propagation is a conventional reference; p-bit LDPC decoding itself is not claimed as the novelty.Source: §I · p. 1
  • The proposed rule retains each bit’s real-valued saturated response and additively reuses it before stochastic readout to test whether bit-local temporal state improves parallel LDPC inference.The contribution is distinguished from temporal memory or same-bit feedback in general; matched controls test amplitude, averaging, identity, and binary inertia.Source: §I · p. 1

A. Factorized objective and instantaneous response

  • The factorized objective is linear in any one spin when the others are fixed, and its derivative supplies that spin’s signed local drive.The derivative is the coefficient of the spin in the multilinear extension, not a restriction to pairwise couplings.Source: A. Factorized objective and… · p. 2

B. Parallel update with stochastic activation

The decoder evaluates candidate fields synchronously while independently deciding whether each p-bit updates or holds. Temporal-state variants then alter the response used for stochastic readout, distinguishing additive reinforcement from bounded filtering and causal controls.

B. Parallel update with stochastic activation

  • All local fields and response candidates use the same pre-update state, making the candidate step Jacobi, while each bit independently holds or activates each cycle.The hold probability is fixed within a run and independent of the annealing schedule; partial activation limits simultaneous commitment and is common to the compared algorithms.Source: B. Parallel update with stochastic… · p. 2

C. Additive and finite-response temporal state

  • Additive memory is reinforcement over [−(1 + λ), 1 + λ], and λ = 0 recovers memoryless pSA; finite-response memory instead stays in [−1, 1].The finite-response state relaxes only on activated updates and attenuates new information when ρ > 0.Source: C. Additive and finite-response… · p. 2
  • The finite-response rule is a sampled single-pole response without a unique physical time-constant assignment for the stochastically gated process.When ungated, ρ = exp(−∆t/τ); with stochastic holding, relaxation occurs only on activated updates.Source: C. Additive and finite-response… · p. 2
  • Normalized memory tests averaging, gain-only tests range without history, and shuffled memory tests whether a distribution-matched but bit-misaligned response suffices.The controls share the Jacobi field evaluation and stochastic activation; a held bit retains both its binary state and stored response.Source: C. Additive and finite-response… · pp. 2–3

D. Causal controls

The causal controls separate temporal averaging, expanded range, bit identity, and the kind of state retained. The LDPC framework evaluates decoding against the transmitted word, where parity validity and channel evidence play distinct roles.

D. Causal controls

  • The controls distinguish expanded deterministic range, temporal averaging, bit identity, and the retained quantity; binary-state self-feedback uses the same bit’s pre-update binary state instead of a separate real-valued response.It is inspired by PIMI but is not a reproduction because the noise, schedule, activation, and workload differ.Source: D. Causal controls · p. 3
  • Both memory rules add one response state per p-bit and O(N) arithmetic, while sparse local-field evaluation remains O(|E|) per cycle.These operation counts do not establish energy, area, latency, or device-efficiency performance.Source: D. Causal controls · p. 3

III. LDPC DECODING AND NUMERICAL FRAMEWORK

  • A zero-syndrome codeword can still be wrong: the channel term favors the transmitted word among valid codewords.The parity term rewards zero syndrome, while the expected channel drive favors the BPSK-consistent bit value.Source: §III · p. 4
  • The primary endpoints are BER and FER relative to the transmitted word at Eb/N0 ∈ {2.0, 2.5, 3.0} dB, for tested sizes N = 96, 192, and 288.Representative matrices are one fixed matrix per size; independent code realizations refer to distinct transfer-study matrices, not finite-size scaling.Source: §III · p. 4

A. Numerical-study designs

Across representative random-regular (3, 6) LDPC codes, additive λ-pSA reduces pooled BER relative to independently optimized pSA. The study then compares numerical designs and controls to examine the source and scope of that advantage.

A. Numerical-study designs

  • Benchmark and control trajectories start from all-zero bit and response states, while the correct-start intervention uses the transmitted codeword and zero response state.Readout may use the final state, post-burn-in bitwise majority, or a visited state selected by syndrome weight and channel score.Source: A. Numerical-study designs · p. 4

IV. REPRESENTATIVE LDPC PERFORMANCE

  • Pooled BER fell by 33.5%, 74.8%, and 81.8% versus independently optimized pSA at N = 96, 192, and 288, respectively.The comparison averages equally over 2.0, 2.5, and 3.0 dB; modes were independently optimized.Source: §IV · p. 4, Table II · p. 5
  • Representative-matrix results alone do not identify the responsible dynamical ingredient or establish transfer to new codes.Matched controls, trajectory analysis, and cross-code tests address those questions.Source: §IV · p. 4

V. CAUSAL CONTROLS OF ADDITIVE TEMPORAL REINFORCEMENT

  • With additive-optimized nonmemory parameters fixed, the study evaluates response rules using 10,000 paired-seed trials per SNR and condition.The selected λ values are 0.95, 0.95, and 0.90 for N = 96, 192, and 288.Source: §V · p. 4
  • At 2.5 dB, shuffled-memory BER exceeds additive BER at N = 96, 192, and 288 despite preserving the stored response's mean and second moment.The shuffle makes each bit's used response nearly uncorrelated with its own preceding response.Source: §V · p. 5
  • At N = 192, tuned binary feedback still has BER 0.21152 at κ = 0.25, exceeding additive by 0.204155; score-selected κ = 0.50 freezes in a wrong valid codeword.The feedback coefficient is swept separately to give binary feedback its best-performance comparison.Source: §V · p. 5
  • Gain-only, normalized averaging, shuffled history, and same-bit binary inertia do not individually reproduce the full saturated-response benefit.Binary feedback nevertheless captures a substantial part of the improvement over matched pSA at N = 288.Source: §V · p. 5

VI. DYNAMICAL MECHANISM IN LDPC DECODING

In the tested LDPC settings, additive response memory is associated with stronger acquisition of the transmitted codeword and post-acquisition stability. Fixed additive packages also outperform pSA-specific packages across the 30 independent code realizations.

A. Acquisition from the default initialization

  • 95.5% and 95.0% of additive trajectories acquired the transmitted word at N = 192 and 288, versus 0% for matched pSA, shuffled memory, and gain-only under the default-start protocol.From random-spin and channel-hard-decision starts, additive acquisition was 96% at N = 192 and 96% and 97% at N = 288; the compared controls were 0% in those cells.Source: A. Acquisition from the default… · pp. 5–6
  • At 2.5 dB, late correct-to-incorrect back-flip rates were 3.63 × 10^-4 and 1.70 × 10^-4 for additive dynamics at N = 192 and 288, versus 0.144 and 0.175 for matched pSA.The bit-aligned retained-response/selected-source product stayed about 0.996, while it collapsed toward zero after shuffling.Source: A. Acquisition from the default… · p. 6

B. Post-acquisition stability and a wrong-codeword counterexample

  • After default-start acquisition, additive trajectories had mean subsequent residence fraction 0.99998 at both N = 192 and 288; in a separate correct-start intervention, residence was 0.9176 and 0.9020.Matched pSA and shuffled memory escaped immediately in the correct-start intervention; conditional on additive escape, return probabilities were approximately 0.959 and 0.963.Source: B. Post-acquisition stability and a… · pp. 6–7
  • Post-acquisition stability is size dependent: at N = 96, additive dynamics could escape more often yet return faster than matched pSA, so the transferable mechanism is not simply escape suppression.Across new N = 96 codes, additive return times spanned roughly 49–84 cycles, compared with 170–349 cycles for the matched pSA ablation.Source: B. Post-acquisition stability and a… · p. 7
  • Low syndrome or low activity alone did not ensure correct decoding: gain-only and the score-selected binary-state control could freeze in wrong valid codewords, while binary feedback at N = 288 acquired the transmitted word in 93% of trajectories.At N = 288, binary feedback decoded BER was 0.01054, above the additive value 0.00595.Source: B. Post-acquisition stability and a… · p. 7

VII. ROBUSTNESS ACROSS INDEPENDENT LDPC CODE REALIZATIONS

  • 35.46%, 76.61%, and 81.10% were the code-level median relative BER reductions at N = 96, 192, and 288; additive packages had lower BER and FER on all 30 codes.Packages were transferred without per-code retuning, and the reductions are relative to the pSA-specific fixed-transfer package.Source: §VII · p. 7, Figure 5 · p. 9
  • Mechanism sampling on nine independent matrices found higher correct-basin acquisition and lower late correct-to-incorrect back-flip rates for additive dynamics in all nine comparisons.The nine matrices comprised three per block length.Source: §VII · p. 7

VIII. DISCUSSION

The discussion argues that additive saturated-response memory changes parallel p-bit LDPC decoding through the retained quantity and its reinjection rule. Its evidence supports benefits within tested settings, while leaving broader generalization and hardware realization unresolved.

VIII. DISCUSSION

  • The improvement has two stages: greater acquisition of the channel-consistent valid-codeword basin, followed by greater stability near that basin or rapid return after departure.Zero syndrome and low activity can also mark a valid wrong-codeword freeze, so success requires channel consistency as well as parity feasibility.Source: §VIII · p. 9
  • Because channel evidence is resampled each cycle, the observed benefit may include temporal integration of channel information as well as stabilization of parallel parity dynamics.The contributions were not separately isolated; separation under a deterministic channel field remains open.Source: §VIII · p. 9
  • Unlike approaches retaining binary state, output-derived adaptive state, or configuration history, this rule retains and reinjects the saturated response itself [14, 15] [12].The discussion attributes computational function to the identity of the retained quantity and its reinjection rule.Source: §VIII · p. 9
  • The fixed-package advantage transfers to independently generated random-regular (3, 6) matrices, but the study does not establish universality across degree distributions, channels, schedules, or block lengths.Initialization robustness was tested on representative N = 192 and 288 matrices at 2.5 dB, with random and channel-hard-decision starts.Source: §VIII · p. 9
  • The results do not demonstrate a circuit realization or energy, area, or speed benefit, and implementation details such as precision, timing, and variability remain dependent on the design.The authors do not identify λ or ρ with a unique device parameter.Source: §VIII · pp. 9–10
  • Supplemental MAX-CUT and random 2-SAT controls do not support a generic advantage of temporal state under the tested conditions.The MAX-CUT results are co-optimized outcomes, whereas the 2-SAT study is a matched-kw finite-response-ρ control.Source: §VIII · p. 10

IX. CONCLUSION

  • Fixed additive packages outperform separately optimized pSA-specific packages across all 30 independent code realizations, while matched λ = 0 effects separately support response-rule causality.Binary self-feedback can contribute but does not reproduce the saturated-response benefit.Source: §IX · p. 10

DATA AVAILABILITY STATEMENT

  • The simulation code, processed data, parameter files, metadata, and figure-reproduction scripts are publicly available through the cited GitHub repository and Zenodo archive.The records include binary-state-control, fixed-transfer, and initialization-robustness validations.Source: DATA AVAILABILITY STATEMENT · p. 10

Appendix A: LDPC sign convention and stochastic channel field

The appendix specifies the LDPC sign and channel conventions, distinguishes codeword validity from decoding correctness, and documents parameter, readout, and statistical definitions. It also describes initialization conditions and acquisition metrics used in the studies.

Appendix A: LDPC sign convention and stochastic channel field

  • With check degree 6, even parity corresponds to a check-product of +1, while the channel field favors the BPSK-consistent bit value.The implementation stores binary bits and maps them to spins as x_i = 2b_i − 1; a positive BPSK sample supports binary zero.Source: Appendix A · p. 11
  • A valid state satisfies every check, but validity alone does not identify the transmitted word; channel consistency distinguishes among valid codewords.The appendix defines correctness as also matching the transmitted word.Source: Appendix A · p. 11

Appendix B: Parameter provenance and readout

  • The benchmark and default-initialization kernels start from all-zero bits and zero response state, whereas robustness and correct-start interventions use distinct bit initializations.Random and channel-hard-decision starts test initialization robustness; the correct-start challenge begins at the transmitted word with zero response memory, and these interventions are not pooled.Source: Appendix B · p. 11
  • Recorded LDPC readouts include the final state, post-burn-in bitwise majority, and a best-visited-state rule that prioritizes syndrome weight and then channel score.The selected readout, burn-in, and sampling window are part of each parameter set.Source: Appendix B · p. 11

Appendix C: Statistical definitions

  • Statistical procedures pair variants through common seed streams, use seed-batch intervals for matched studies, and bootstrap independent code realizations for cross-code transfer.Percentile intervals use 20,000 bootstrap resamples.Source: Appendix C · pp. 11–12

Appendix D: Acquisition and stability metrics

  • Correct-reach probability is the fraction of trajectories that visit the transmitted codeword within the simulation horizon.The acquisition metric defines the first hit as the first cycle at which the trajectory reaches that word.Source: Appendix D · p. 12

affects I(t)

The supplemental methods define response-hold conventions, benchmark optimization, and matched controls for LDPC decoding. The matched controls report positive paired control-minus-additive BER differences across all nine comparisons.

S2. INDEPENDENTLY OPTIMIZED REPRESENTATIVE LDPC BENCHMARK

step-0.02 grid within ±0.15 of the confirmed coefficient, subject to the mode bounds. Every

B. BP reference decoder

The matched causal-control study fixes the size-specific additive-optimized parameter

  • Paired control-minus-additive BER intervals are positive in all nine comparisons, and setting λ = 0 makes all four kernels trajectory-identical under common random streams.The additive coefficient sweep found no size with continuing improvement at its enlarged upper boundary.Source: The matched causal-control study fixes… · p. 19

S4. BINARY-STATE SELF-FEEDBACK CONTROL

trial-per-SNR evaluation used for publication-facing controls. The prespecified sweep score

  • The prespecified binary-feedback score selects κ = 0.25, 0.50, and 0.75 for N = 96, 192, and 288, respectively.The score pools BER, FER, and nonzero-syndrome fraction over the three SNRs; the selected κ is the minimum on the declared grid.Source: trial-per-SNR evaluation used for… · p. 21

At N = 96 and 288, the score-selected values are also the BER- and FER-minimizing

Binary-state feedback can improve acquisition at N = 288 but does not reproduce the additive rule’s BER advantage, while at N = 192 it can freeze in a valid but wrong codeword. Response-alignment diagnostics also show that shuffling delayed responses worsens BER relative to additive dynamics.

shows a valid wrong-codeword freeze. To avoid presenting that pathological score selection

  • At N = 192, the publication-facing binary comparison uses a BER-favorable κ = 0.25 sensitivity rather than the score-selected κ = 0.50.The substitution is unique to N = 192 and does not change the archived score selection.Source: shows a valid wrong-codeword freeze.… · p. 22

N = 288. Their binary BERs are 0.49951, 0.49981, and 0.34739, compared with additive

  • The additive BERs were 0.01576, 0.00749, and 0.00740, while matched-pSA BERs were 0.03581, 0.28087, and 0.32236; amplitude matching does not reproduce the additive rule.Source: N = 288. Their binary BERs are… · p. 22

substituted for the best-κ estimand in Table S5

  • At N = 192, score-selected binary feedback acquires the transmitted word in 0% of 200 trajectories and has decoded BER 0.497 despite reaching zero syndrome with negligible motion.At N = 288, κ = 0.75 acquires the word in 93% of trajectories and retains it, but decoded BER 0.01054 remains above additive BER 0.00595.Source: substituted for the best-κ estimand in… · pp. 22–23

additive/normalized rules and s(t)

  • At 2.5 dB, the reported paired shuffled-minus-additive BER differences include 0.0800 and 0.3763; the cited passage does not provide a third value or associate these values with N.Their corresponding 95% intervals are [0.0747, 0.0853] and [0.3704, 0.3822].Source: additive/normalized rules and s(t) · p. 24

S5. RESPONSE ALIGNMENT AND TRAJECTORY DIAGNOSTICS

  • The response diagnostics distinguish late correct-to-incorrect back-flip rate from retained-response/selected-source product, which includes held cycles.The selected-source product is not applicable to pSA or gain-only because they supply no delayed response source.Source: additive/normalized rules and s(t) · p. 24, §0.3913 · p. 25

S6. ACQUISITION AND POST-ACQUISITION STABILITY

The trajectory studies separate default-start acquisition from correct-start retention and test whether acquisition depends on starting at the all-zero codeword. Under the tested alternative starts, additive dynamics retain a large acquisition advantage over matched pSA.

Default initialization and correct-start retention are distinct interventions. The default

  • The correct-start intervention initializes bits at the transmitted word but response memory at zero, so it measures retention before matching history has accumulated.It is distinct from the default-start acquisition study, which begins at the implementation’s all-zero state, generally not the transmitted word.Source: Default initialization and… · p. 25
  • In the correct-start challenge, additive residence is 0.9176 at N = 192 and 0.9020 at N = 288, while pSA and shuffled memory leave immediately.Gain-only residence is one at N = 192 but zero at N = 288, showing that residence without channel-consistent acquisition is not a transferable success mechanism.Source: Default initialization and… · pp. 25–26
  • From default initialization, gain-only at N = 192 freezes at zero syndrome and zero late flip rate but has late BER 0.497 and normalized channel alignment 0.0074.The corresponding additive values are late BER 0.0066 and normalized channel alignment 0.695.Source: Default initialization and… · p. 27

A. Initialization robustness

  • At N = 192, additive acquisition exceeded matched pSA by 96.0 percentage points in both comparisons; at N = 288, differences were 96.0 and 97.0 points, respectively.Source: A. Initialization robustness · p. 29

Across the four alternative-initialization cells, additive late instantaneous BER is 0.0038–

  • The initialization-robustness results apply to the tested representative matrices, block lengths, 2.5-dB channel, and random or channel-hard-decision starts; broader dependence is not established.Matched pSA does not reach the correct word in any alternative-start cell, so its conditional first-passage-time difference is undefined.Source: A. Initialization robustness · p. 28, Across the four… · pp. 30–31

The fixed-transfer evaluation has three distinct arms. The additive and pSA-specific

Fixed-transfer additive packages outperform pSA-specific packages across the 30 code realizations, while matched ablations and trajectory analyses probe the response rule and its dynamics. Cross-problem controls delimit what can be inferred beyond LDPC decoding.

The fixed-transfer evaluation has three distinct arms. The additive and pSA-specific

  • The fixed-transfer comparison preserves each package's selected readout and other fields; the matched λ = 0 ablation changes only the response coefficient.The additive and pSA-specific packages were selected independently on representative matrices, then transferred intact to the same C00–C09 matrices.Source: The fixed-transfer evaluation has… · p. 31

code instances, SNRs, stochastic seed design, channel generation, and initialization. Main

Code Matched BER Add. BER Red. (%) Matched FER Add. FER

S8. CROSS-PROBLEM BOUNDARY CONTROLS

  • The cross-problem controls do not establish a generic endpoint advantage from temporal state or support the LDPC mechanism.MAX-CUT outcomes are co-optimized, while random 2-SAT uses a matched-kw finite-response control.Source: S8. CROSS-PROBLEM BOUNDARY CONTROLS · p. 36

The implementation minimizes

The broader N = 500 structural campaign covers 14 density/topology conditions, ten

  • Across graph seeds and 100 trials per graph, only 5 of 14 mean-quality winners use a nonzero response coefficient, while apparent mean normalized-cut gains remain below 0.2%.The authors characterize the evidence as marginal or conditional.Source: The broader N = 500 structural… · p. 36

B. Random 2-SAT matched-kw control

E2SAT/kw is exactly the unsatisfied-clause count.

The matched-kw random 2-SAT control does not support a positive response-memory effect: its predeclared contrast worsens the unsatisfied-clause rate. The section also distinguishes that control from exploratory co-optimization and documents validation and reproducibility checks.

E2SAT/kw is exactly the unsatisfied-clause count.

  • The initial exploratory study co-selected kw and ρ, so its apparent density trend cannot establish an independent response-memory effect.The authors identify the matched-kw control as superseding that trend.Source: E2SAT/kw is exactly the… · p. 38, §17 · p. 39

the displayed precision, excluding a hidden improvement at the earlier kw search boundary.

17 unsatisfiable), 100 paired trials per formula and cell, and 3000 cycles. It evaluates kw ∈

  • In the matched-kw response surface, ρ = 0 is best in every clause-weight row, and the global minimum is kw = 8, ρ = 0.The reported rate degrades approximately monotonically with nonzero ρ for every kw.Source: §17 · p. 39, §0.00211222 · p. 40, Table S13 · p. 41

0.00211222 to 0.00215167: treatment minus control is +3.944×10−5 with formula-bootstrap

standardized effect dz = 0.519. Terminal unsatisfied rate worsens by 11.35%.

  • Among unsatisfiable formulas, selected unsatisfied-clause rate increases by 4.804 × 10−5, with interval [0.882, 8.627] × 10−5.The passage states that neither satisfiability stratum provides positive response-memory evidence.Source: standardized effect dz = 0.519.… · pp. 40–41

S9. VALIDATION AND INVARIANT TESTS

  • Validation checks confirm trajectory identity at λ = 0, ρ = 0, and κ = 0 under matched parameters and shared random streams.The checks also verify shuffled mappings and cross-code matrices, among other implementation invariants.Source: S9. VALIDATION AND INVARIANT TESTS · pp. 41–42

11. Reported cross-code intervals resample codes, initialization-acquisition intervals use

  • Reported cross-code intervals resample codes, initialization-acquisition intervals use ten paired seed-batch means, and matched-kw intervals resample formulas.The method does not treat bits as independent instances for those claims.Source: §11 · p. 42

For implementation provenance, the optimized LDPC Numba path for the memoryless,

  • The public release documents implementation paths and provides scripts, dependencies, parameters, seeds, metadata, checksums, and reproduction commands.The release also includes binary-state-control, three-arm transfer, and initialization-validation records.Source: For implementation provenance, the… · p. 42
Loading 2609.12389v1…