Plain-language summary: what changed across versions

This record has gone through three stages. The short version is:

• v1.1 made a broad claim.

• v1.2 narrowed the claim after new idle-control experiments.

• v1.3 corrects the record after checking which results can actually be reproduced from the deposited files.

The current conclusion is not that the whole project failed. The current conclusion is more specific:

We still have a reproducible spatial anomaly in one Y⊗Z stabilizer experiment, and we also have a clean negative-control result showing no generic idle bit-flip correlation across three IBM Heron-class systems. However, the earlier temporal-memory and reset/reprepare claims cannot currently be reproduced from the deposited data and code, so those claims are withdrawn pending a proper reproducible reanalysis.

v1.1 — What did we do?

In v1.1, we reported a three-part experiment on IBM’s ibm_fez quantum processor:

1. a spatial test, asking whether groups of nearby qubits behaved independently;

2. a temporal test, asking whether one measurement influenced a later measurement;

3. a reset/reprepare test, asking whether memory could survive even after the qubits were reset and prepared again.

At the time, the results appeared to point in the same direction: quantum errors seemed spatially structured, time-dependent, and possibly influenced by the surrounding device environment.

v1.1 — What did we claim then?

The original description used strong language, including phrases like “complete characterization,” “unassailable proof,” “smoking gun,” and “standard QEC assumptions now experimentally disproven.”

That language is now superseded.

v1.1 — What still stands?

The spatial test still stands in a narrower form. When re-run from the deposited data and code, the Phase 1 spatial Y⊗Z parity-consistency anomaly reproduces.

v1.1 — What no longer stands?

The temporal-memory and reset/reprepare claims do not currently stand as published claims. They are not proven false, but the deposited files do not contain a complete reproducible path from the raw IBM job results to the reported conditional probabilities and σ-values.

Therefore, those claims are withdrawn pending reproducible reanalysis.

v1.2 — What did we do?

In v1.2, we added a new control experiment using simple idle circuits.

In plain language, this asked:

If we prepare qubits in the simplest possible way, let them sit idle, and then measure them, do they fail together?

This was tested across three IBM Heron-class systems:

• ibm_fez

• ibm_kingston

• ibm_marrakesh

The experiment looked for pairwise bit-flip correlation during idle time.

v1.2 — What did we find?

The result was a clean null:

No tested qubit pair showed statistically significant idle bit-flip correlation across the three devices.

This means the earlier spatial anomaly is not explained by a large, generic background effect where idle qubits simply flip together all the time.

v1.2 — What did we stop claiming?

v1.2 already narrowed the original story. It moved away from the idea that Heron-class hardware has a universal, always-on correlated-noise problem.

Instead, the better interpretation became:

If a memory/correlation effect exists, it is likely protocol-specific or observable-specific, not generic idle noise.

v1.3 — What did we do?

In v1.3, we performed a reproducibility correction.

We asked a simple question:

Which results can be regenerated from the deposited data and code, and which cannot?

The answer split the record into two parts.

v1.3 — What still stands?

Two results stand.

First, the Phase 1 spatial Y⊗Z parity anomaly is retained. It reproduces from the deposited Phase 1 IBM job result and analysis script. After correcting for the six modules tested, the strongest module still survives strict multiple-comparison correction.

Second, the three-backend idle-control null is retained. The idle-control pipeline is deposited and reproducible, and it shows no detectable pairwise ambient idle bit-flip correlation at the current sensitivity scale.

v1.3 — What is withdrawn?

The Phase 2 temporal-memory claim and the Phase 3B reset/reprepare claim are withdrawn pending reproducible reanalysis.

This is not because they have been proven false. It is because the deposited artifacts do not currently show how to compute the published numbers from the raw data. The executed circuit structure was also not deposited in enough detail to verify exactly what circuit was run.

v1.3 — What supersedes the prior versions?

The current v1.3 interpretation supersedes the broad claims in v1.1 and v1.2.

The surviving claim is:

A reproducible, protocol-specific spatial parity anomaly was observed in the Y⊗Z stabilizer experiment on ibm_fez. However, broad temporal/environmental-memory claims are withdrawn pending reproducible reanalysis, and controlled idle experiments across three Heron-class systems do not show generic pairwise idle bit-flip correlation.

In simpler terms:

We found one reproducible spatial anomaly. We did not find generic idle correlation. And we withdrew the earlier temporal-memory claims until they can be reproduced cleanly.

Why this matters

This correction is not a retraction of the whole project. It is a narrowing of the claim to what the deposited data can support.

The lesson is important:

In quantum-hardware experiments, it is not enough to save the measurement counts. A complete record must also preserve the exact circuit, the transpiled circuit, the qubit layout, the backend, the calibration context, and the analysis code that derives every reported number.

This v1.3 version is therefore both a corrected scientific record and a process improvement for future experiments.

 

Reproducible Spatial Parity Anomaly in Heron-Class Processors, with Idle-Correlation Controls and Temporal-Claim Withdrawal (v1.3)

Status of this version

This version is a reproducibility correction. On systematic reanalysis of the deposited artifacts, the claims separate into three groups:

• Retained (reproducible): the Phase 1 spatial Y⊗Z parity-consistency anomaly.

• Retained (reproducible): the follow-up three-backend idle-correlation null.

• Withdrawn pending reproducible reanalysis: the Phase 2 temporal-memory and Phase 3B reset/reprepare claims, including the earlier 30 µs convergence narrative built on them.

The underlying raw job results are unchanged and remain openly verifiable. What changed is the assessment of which reported numbers can be derived from those results using deposited code.

Why the temporal/reprepare claims are withdrawn

The Phase 2 and Phase 3B results are withdrawn — not because they are shown to be false, but because they cannot currently be reproduced from the deposited artifacts.

Specifically:

1. No analysis pipeline was located that derives the reported conditional-dependence probabilities, such as the Phase 3B values P(2nd=0 | 1st=0)=0.4707 and P(2nd=0 | 1st=1)=0.5183 at 30 µs, from the raw job results. The values appear only in hardcoded summary form in a display script.

2. Direct recomputation of the conditional dependence from the raw measurement register yields marginals inconsistent with the published values, approximately 0.57 rather than the reported 0.47–0.52, indicating that the published numbers were produced by a transformation not present in the deposited code.

3. The executed circuit structure was not deposited: circuit_metadata is empty, and the metadata records only protocol, center, data qubits, delay, and quality score. Therefore, the run cannot be verified against the available circuit-construction code.

4. Even under the most favorable statistical reading, the effect appeared at only 2 of 4 delays and was reported as single-test σ-values, without permutation nulls or multiple-comparison correction.

Stated precisely: the deposited artifacts do not contain a complete provenance chain from executed circuit → raw counts → reported conditional probabilities and σ-values.

The Phase 2/3B claims are therefore withdrawn pending a documented, reproducible reanalysis.

Retained Result 1 — Phase 1 spatial Y⊗Z parity anomaly

The Phase 1 spatial analysis does reproduce. The deposited script analyze_yz_syndrome_sweep.py loads the Phase 1 job result job-d61v0lao8gvs73f1gutg-result.json, decodes the measurement data, and computes the parity-consistency test YZ₀₁·YZ₁₂ = YZ₀₂ across six four-qubit modules.

Re-running the script reproduces the published per-module values. A source-code search found no hardcoded instances of the reproduced values in the analysis script.

Multiple-comparison correction across six modules

Module	Qubits	σ (single-test)	p (2-sided)	p Bonferroni (×6)	q (BH-FDR)	Survives
0	[0,1,2]	4.86	1.2×10⁻⁶	7.0×10⁻⁶	7.0×10⁻⁶	Bonferroni + FDR
2	[8,9,10]	3.76	1.7×10⁻⁴	1.0×10⁻³	5.1×10⁻⁴	Bonferroni + FDR
1	[4,5,6]	2.96	3.1×10⁻³	1.9×10⁻²	6.2×10⁻³	Bonferroni + FDR
5	[17,27,26]	2.35	1.9×10⁻²	1.1×10⁻¹	2.8×10⁻²	FDR only
4	[16,23,22]	1.89	5.9×10⁻²	3.5×10⁻¹	7.1×10⁻²	ns
3	[12,13,14]	1.05	2.9×10⁻¹	1.0	2.9×10⁻¹	ns

Module 0 survives strict Bonferroni correction by a wide margin (p = 7×10⁻⁶ after correction). Modules 2 and 1 also survive Bonferroni correction. The spatial anomaly is therefore robust to multiple-comparison correction across the six scanned modules.

Caveat

Module 0 — the strongest violation — also has the lowest primary measurement fidelity of all six modules (YZ₀₁ = 0.553, versus 0.74–0.83 for the others). A parity-consistency violation co-located with degraded primary measurement warrants caution: the anomaly may reflect protocol-specific measurement degradation or observable-specific error structure on qubits 0/1 rather than a general spatial non-Markovian mechanism.

The ancilla-only control (0.962) argues against pure readout error but does not fully exclude the confound. Follow-up with improved readout calibration, repeated module selection, and full circuit provenance is needed before the spatial anomaly is interpreted as definitively non-Markovian.

Appropriately scoped claim

A reproducible Y⊗Z parity-consistency anomaly is observed on selected ibm_fez modules, strongest on Module 0 (4.86σ single-test; corrected p = 7×10⁻⁶), surviving multiple-comparison correction, with apparent topology dependence. This is a protocol- and observable-specific anomaly, not a demonstration of universal non-Markovian error behavior.

Retained Result 2 — three-backend idle-correlation null

Controlled Z-basis idle/identity experiments were run across ibm_marrakesh, ibm_kingston, and ibm_fez using 12-qubit contiguous heavy-hex patches, 8192 shots, and idle delays spanning 0–120 µs. ibm_marrakesh used τ = {0,10,30,60,120} µs; ibm_kingston and ibm_fez used τ = {0,30,120} µs.

No qubit pair on any tested backend showed residual bit-flip mutual information surviving a permutation null at 99% family-wise significance, with Bonferroni correction over 66 pairs. The detection floor was approximately 4×10⁻⁴ bits MI. The null held across a roughly 2× spread in baseline error rate.

The full pipeline — submission, fetch, analysis, and permutation test — is deposited and reproducible.

This is a strong controlled negative: pairwise ambient idle bit-flip correlation was not detected at the ~10⁻⁴-bit MI scale across the tested hardware. It does not falsify the broader possibility of protocol-induced, phase-like, active-gate, or higher-order correlated effects; it constrains the interpretation of the retained spatial anomaly.

Combined interpretation

The project is not invalidated; it is sharpened and bifurcated:

• There is reproducible evidence of a protocol-specific spatial parity anomaly in Y⊗Z stabilizer measurements on ibm_fez Phase 1 data.

• There is no detected evidence of generic pairwise ambient bit-flip correlation in controlled Z-basis idle experiments across three Heron-class devices.

• The temporal/environmental memory claims are withdrawn pending reproducible reanalysis.

Surviving claim:

Reproducible protocol-specific spatial parity anomaly + clean multi-device idle null; temporal/reprepare claims withdrawn pending a documented, reproducible pipeline.

In plain language: one part of the original experiment still holds up, and two parts do not currently hold up.

The part that survives is the spatial result: one Y⊗Z stabilizer experiment on ibm_fez showed a reproducible anomaly in how certain groups of qubits behaved together. That result can be regenerated from the deposited data and code, and the strongest module still survives multiple-comparison correction.

The new idle-control experiments also hold up: when simple idle circuits were run on three Heron-class systems — ibm_fez, ibm_kingston, and ibm_marrakesh — no generic pairwise idle bit-flip correlation was detected. In other words, the hardware did not show an obvious background pattern where idle qubits simply failed together.

What no longer stands are the earlier temporal-memory and reset/reprepare claims. Those claims are not proven false, but the deposited files do not currently show a complete, reproducible path from the raw IBM results to the published numbers. They are therefore withdrawn until a documented analysis pipeline can reproduce them.

So the corrected conclusion is narrower but stronger: there is a reproducible spatial parity anomaly in a specific Y⊗Z protocol, there is no detected generic idle bit-flip correlation across the tested devices, and the earlier time-memory claims are no longer claimed until they can be reproduced cleanly.

Methodology lessons

This correction was necessary because the original deposit lacked complete provenance discipline. Going forward, every experiment in this line should deposit a complete chain:

1. source circuit,

2. transpiled circuit,

3. final layout,

4. backend name and calibration timestamp,

5. job ID,

6. raw counts,

7. analysis script that loads the raw data and derives every reported number,

8. generated metrics JSON.

Reported significances should include permutation/bootstrap nulls and multiple-comparison correction specified before analysis. Analysis scripts should not contain hardcoded result values except for display labels or explicitly documented reference values.

Note on earlier versions

Earlier versions of this record used absolute language such as “complete characterization,” “unassailable proof,” “now experimentally disproven,” “discovery-level,” and “smoking gun.” Those characterizations are superseded by this version and should not be cited.

Readers citing prior versions should refer to this correction for the current status of each claim.

Data, verification, and licensing

Raw job IDs are unchanged:

• Phase 1: d61v0lao8gvs73f1gutg

• Phase 2: d62h65ns6ggc73fgqee0

• Phase 3A: d62lmg3c4tus73fdkb9g

• Phase 3B: d62lmurc4tus73fdkbo0

Idle-study data, scripts, permutation-test outputs, and correction code are included in this version.

Data are released under CC BY 4.0. Code is released under the MIT License. The Y⊗Z π/4 stabilizer method is covered by U.S. Provisional Patent 63/952,786; this dataset does not disclose enabling detail.

Principal Investigator: Amit Brahmbhatt, Quantum-Clarity LLC.