When Reset Is Not Enough: Evidence for Protocol-Dependent Error Memory in Heron-Class Quantum Processors

Protocol-Dependent Evidence for Spatial and Temporal Error Memory in Heron-Class Processors, with Subsequent Idle-Correlation Controls Overview This dataset reports a multi-protocol experimental campaign investigating whether error behavior on IBM 156-qubit Heron-class superconducting processors departs from a simple independent, memoryless noise model. The original campaign was conducted on ibm_fez and included three active protocols: a spatial Y⊗Z parity-consistency test, a temporal sequential-measurement test, a reset/repreparation test designed to probe whether conditional dependence can persist after full qubit reset. A subsequent follow-up control study tested deliberately minimal Z-basis idle/identity circuits across three Heron-class systems: ibm_fez, ibm_kingston, and ibm_marrakesh. The revised interpretation is intentionally narrower than the original version of this record. The original protocols provide evidence consistent with protocol-dependent spatial and temporal error memory, including a recurring feature near 30 µs. However, the reported significances are single-test σ-values and should be treated as provisional until permutation/bootstrap nulls and multiple-comparison correction are applied. The follow-up idle-correlation study found no detectable pairwise ambient Z-basis bit-flip correlation across three Heron-class devices at the ~10⁻⁴-bit mutual-information scale. This means the original active-protocol signals, if reproduced under corrected statistics, should be interpreted as protocol-dependent or operation-induced memory rather than as evidence for a large generic ambient idle-bit-flip correlation present at all times. In plain terms: this dataset suggests that certain measurement, reset, repreparation, or stabilizer-construction protocols may expose memory effects in the hardware environment. It does not establish that all Heron-class errors are universally correlated or that standard quantum error-correction assumptions are globally invalid. What changed in this version Earlier description text used overly absolute language, including phrases such as “complete characterization,” “unassailable proof,” “definitively detected,” “now experimentally disproven,” and “discovery-level.” That language overstated what the reported statistics can support. This version reframes the record to match the evidence more carefully: The original experiments are described as protocol-dependent evidence, not final proof. Reported σ-values are identified as single-test values. The need for permutation/bootstrap nulls and multiple-comparison correction is stated explicitly. A new three-backend idle-correlation control is incorporated into the interpretation. The central claim is narrowed from “errors are universally non-Markovian” to “specific active protocols may reveal memory effects not visible in simple idle controls.” The underlying original job results and job IDs remain unchanged and publicly verifiable. Experimental campaign summary Platform Original campaign: Backend: ibm_fez Architecture: IBM Heron-class superconducting processor Processor size: 156 qubits Campaign dates: February 4–6, 2026 Total QPU time: approximately 73 seconds Total measurements: approximately 500,000 circuit executions Follow-up idle-control study: Backends: ibm_fez, ibm_kingston, ibm_marrakesh Architecture: IBM Heron-class superconducting processors Processor size: 156 qubits each Protocol: Z-basis idle/identity circuits Patch size: 12 qubits Shots: 8192 per circuit Idle durations: 0–120 µs Analysis: pairwise mutual information with permutation nulls and Bonferroni correction Original protocols on ibm_fez The original three active protocols used a proprietary π/4 Y⊗Z stabilizer construction associated with U.S. Provisional Patent Application No. 63/952,786. This dataset reports experimental results without disclosing enabling implementation detail. Phase 1 — Spatial Y⊗Z parity-consistency test The spatial protocol tested an algebraic parity-triangle consistency relation among Y⊗Z stabilizer observables: YZ₀₁ · YZ₁₂ = YZ₀₂ (mod 2) Under the protocol’s independence model, the directly measured YZ₀₂ value is expected to agree with the value inferred from YZ₀₁ and YZ₁₂. Deviations from this relation may indicate spatially structured or topology-dependent error behavior, subject to appropriate controls and statistical correction. Protocol summary: Six four-qubit modules tested across chip topology 36 circuits total 8192 shots per circuit Includes syndrome and control circuits Designed to compare directly measured and inferred parity structure Module Qubits Measured YZ₀₂ Expected YZ₀₂ Deviation Reported σ* Interpretation 0 [0,1,2] 0.4725 0.4994 2.68% 4.86 strongest single spatial deviation 2 [8,9,10] 0.5339 0.5132 2.07% 3.76 significant single-module deviation 1 [4,5,6] 0.5187 0.5023 1.64% 2.96 marginal 5 [17,27,26] 0.4836 0.4966 1.30% 2.35 marginal 4 [16,23,22] 0.5066 0.4961 1.05% 1.89 consistent within reported uncertainty 3 [12,13,14] 0.5125 0.5067 0.58% 1.05 consistent within reported uncertainty *Reported σ-values are single-test values and should be reassessed with correction for the six modules scanned. The strongest single result is Module 0 at 4.86σ before multiple-comparison correction. The module-to-module variation suggests possible topology dependence, but the statistical status should be treated as provisional until reanalyzed with family-wise or false-discovery correction. Phase 2 — Temporal sequential-measurement test The temporal protocol tested whether a second measurement outcome depends on a prior measurement outcome after a controlled delay. Protocol structure: SIGNAL: measure → reset ancilla → delay Δt → measure again ANCILLA-ONLY: ancilla-only control sequence ONE-SHOT: single-measurement timing baseline Delays tested: 0, 2, 10, 30, 50, 150 µs Memory metric: |P(2nd=0 | 1st=0) − P(2nd=0 | 1st=1)| Delay (µs) SIGNAL ANCILLA-ONLY Reported σ* Interpretation 0 0.0113 0.0098 baseline baseline/control level 2 0.0165 — — modest increase 10 0.0092 — — valley 30 0.0362 0.0098 2.8 peak reported temporal feature 50 0.0147 — — decay 150 0.0277 — — suggestive revival *Reported σ-values are single-test values and should be reassessed after correction across delays and protocol variants. The temporal protocol shows its largest reported effect at 30 µs, with a reported 2.8σ single-test significance. This is suggestive but below the 3σ threshold commonly treated as stronger evidence. The apparent non-monotonic pattern is physically interesting but should not be overinterpreted without corrected null analysis. Phase 3B — Reset/repreparation test The reset/repreparation protocol was designed to test whether conditional dependence can persist after full qubit reset and fresh preparation. Protocol structure: measure YZ₀₁ → reset all qubits → fresh preparation from |0000⟩ → measure YZ₀₁ again The key idea is that reset should remove ordinary qubit-state memory. If conditional dependence persists after reset and fresh preparation, one possible interpretation is that memory resides in environmental, control, resonator, or material degrees of freedom rather than in the qubit state itself. Delay (µs) P(2nd=0 | 1st=0) P(2nd=0 | 1st=1) Difference Reported σ* 0 0.5032 0.5039 0.0006 0.0 10 0.4996 0.4857 0.0139 1.0 30 0.4707 0.5183 0.0476 3.6 50 0.4783 0.5115 0.0332 2.4 *Reported σ-values are single-test values and should be reassessed after correction across delays. The 30 µs point is the strongest reset/repreparation result, with a reported 3.6σ single-test significance. The 50 µs point is suggestive at 2.4σ. The 0 µs and 10 µs points do not show comparable significance. This protocol is physically interesting because a reproducible conditional dependence after full reset would suggest that reset does not erase all relevant physical memory in the broader device environment. Possible mechanisms include residual readout photons, TLS defects, cavity modes, local heating, or control-line history. However, because the significance appears only at selected delays and was not originally tested against a permutation/bootstrap null with multiple-comparison correction, the effect should be described as suggestive pending corrected reanalysis, not definitive proof. Follow-up controlled idle-correlation study A follow-up control study was performed to test whether the original active-protocol observations reflect a generic ambient property of Heron hardware. The idle-control protocol was deliberately minimal: prepare |0…0⟩ → idle for τ → measure in Z basis The ideal output is all zeros. Therefore, any measured 1 can be treated as a bit-flip error event, and pairwise mutual information between error events can be used to test for ambient Z-basis bit-flip correlation without designed-state contamination. Setup Backends: ibm_fez, ibm_kingston, ibm_marrakesh Patch size: 12 qubits Patch type: contiguous heavy-hex patch Idle durations: 0–120 µs Shots: 8192 Basis: Z Dynamical decoupling: off Statistic: pairwise mutual information between residual error events Null model: 500 permutations per pair Multiple-comparison correction: Bonferroni over 66 qubit pairs Family-wise significance level: 99% Results Backend τ range Error-rate range Max raw MI Detection floor (99%) Significant pairs ibm_marrakesh 0–120 µs 0.52–0.64% 8.4×10⁻⁴ 3.7×10⁻⁴ 0 ibm_kingston 0–120 µs 0.89–1.07% 5.8×10⁻⁴ 4.2×10⁻⁴ 0 ibm_fez 0–120 µs 0.70–0.77% 4.0×10⁻⁴ 4.2×10⁻⁴ 0 No qubit pair on any tested device showed residual error mutual information surviving the permutation null at 99% family-wise significance. This null result generalized across three Heron-class systems with different baseline error rates. Kingston showed the highest baseline error rate, Marrakesh the lowest, and Fez was intermediate, yet all three returned the same structureless null at approximately the same detection scale. Several raw pairwise MI values approached or exceeded the per-run floor, but none survived multiple-comparison correction. This is the expected behavior of a sound family-wise test rejecting isolated fluctuations under the null. Sharp interpretation The follow-up idle study does not show that “errors are never correlated.” It supports the narrower statement: Controlled Z-basis idle circuits do not reveal detectable pairwise ambient bit-flip error correlation at the ~10⁻⁴-bit mutual-information scale across the tested Heron-class systems. Reconciliation of the original active protocols and the idle-control null The original active protocols and the follow-up idle study measure different observables. The idle study asks: “Do qubits spontaneously flip together while sitting idle in the Z basis?” The reset/repreparation protocol asks: “Does the outcome of a later active measurement depend on an earlier active measurement after reset and fresh preparation?” These are not the same physical question. A superconducting processor can have no detectable ambient pairwise Z-basis idle bit-flip correlation while still exhibiting protocol-induced memory after measurement, reset, repreparation, or stabilizer construction. Reset returns the qubit state to |0⟩, but it does not necessarily reset all surrounding physical degrees of freedom, such as readout resonators, residual photons, TLS defects, cavity modes, local thermal transients, or control-line history. Therefore, the idle null does not automatically invalidate the active-protocol observations. Instead, it sharpens their interpretation. If the active-protocol effects survive corrected reanalysis, then the appropriate conclusion is not that Heron hardware has large generic ambient bit-flip correlations. The stronger and more precise conclusion would be that certain active operations may inject or expose transient environmental memory that persists beyond qubit reset. If the active-protocol effects do not survive corrected reanalysis, then the original observations should be treated as likely statistical or protocol-analysis artifacts. The current dataset therefore motivates a stricter reanalysis rather than a universal claim. Statistical status and citation guidance The original three-phase reported σ-values are single-test significances. They have not yet been subjected to: permutation or bootstrap null distributions, correction across scanned delays, correction across scanned modules, family-wise error control, false-discovery control. The follow-up idle study was analyzed with permutation nulls and Bonferroni family-wise correction and returned null results. Until the active protocols are reanalyzed to the same standard, they should be cited as: “suggestive, protocol-dependent evidence of spatial and temporal error memory” rather than as: “definitive proof,” “unassailable proof,” or “universal disproof of independent error models.” A corrected reanalysis of the openly available Phase 2 and Phase 3 job results is planned. If the active-protocol features survive stricter null testing, the most defensible interpretation will be protocol-induced environmental memory rather than generic ambient idle correlation. Implications for quantum error correction The appropriate implication is not that all standard quantum error-correction assumptions are universally false. The narrower implication is that certain measurement, reset, repreparation, or stabilizer-extraction protocols may produce temporally structured or topology-dependent error behavior. If confirmed under corrected statistics, such behavior would matter for repeated syndrome extraction, decoder confidence weighting, hardware-aware scheduling, and operation-aware mitigation. The follow-up idle-control result suggests that decoders do not need to assume large ambient Z-basis bit-flip correlation on the tested Heron-class devices. However, decoders and runtime systems may still need to account for operation-induced correlations, phase/dephasing memory, measurement-chain effects, or higher-order multi-qubit correlations. This distinction is important: Ambient idle bit-flip correlation: not detected in the follow-up control study. Protocol-induced memory: suggested by the original active protocols, pending corrected reanalysis. Phase/dephasing correlation: not yet tested by the Z-basis idle control. Active-gate crosstalk: not resolved by the idle study and remains a possible location for correlated effects. Data availability All original job results remain publicly verifiable using IBM Quantum job IDs. Original campaign: Phase 1 spatial protocol: d61v0lao8gvs73f1gutg Phase 2 temporal sequential protocol: d62h65ns6ggc73fgqee0 Phase 3A echo protocol: d62lmg3c4tus73fdkb9g Phase 3B reset/repreparation protocol: d62lmurc4tus73fdkbo0 Dataset contents include raw job-result JSON files, metadata files, README documentation, citation metadata, and licensing information. Follow-up idle-control job IDs and analysis outputs will accompany the planned corrected reanalysis release. Licensing and IP note Data are released under Creative Commons Attribution 4.0 International (CC BY 4.0). Analysis scripts are released under the MIT License where provided. The Y⊗Z stabilizer preparation method using π/4 rotations is associated with U.S. Provisional Patent Application No. 63/952,786. This dataset reports experimental results but does not disclose enabling implementation details. Version history v1.0 — February 5, 2026 Initial Phase 1 spatial syndrome sweep. Previous DOI: 10.5281/zenodo.18498540 v1.1 — February 6, 2026 Original three-phase campaign: Phase 1: spatial Y⊗Z parity-consistency test Phase 2: temporal sequential-measurement test Phase 3: reset/repreparation test DOI: 10.5281/zenodo.18501679 v1.2 — May 2026 Revised interpretation and description: Reframed claims as protocol-dependent evidence rather than definitive proof. Added statistical caveats for single-test σ-values. Integrated follow-up three-backend Z-basis idle-correlation controls. Clarified that the idle controls found no detectable ambient pairwise bit-flip correlation at the ~10⁻⁴-bit MI scale. Narrowed the scientific interpretation to possible protocol-induced memory pending corrected reanalysis. Citation @dataset{brahmbhatt_protocol_dependent_error_memory_2026,author = {{Amit Brahmbhatt}},title = {{When Reset Is Not Enough: Evidence for Protocol-Dependent Error Memory in Heron-Class Quantum Processors}},year = 2026,month = may,publisher = {Zenodo},version = {1.2},doi = {10.5281/zenodo.20467541},url = {https://doi.org/10.5281/zenodo.20467541}}