Protected-Plane Coherence Lifetime and Stabilizer Refresh on IBM Kingston: A Three-Curve Y⊗Z Hold-Time Study

Y⊗Z Protected Plane Hold-Time Experiment: Three-Curve Discrimination on IBM Kingston

Author: Amit BrahmbhattOrganisation: Quantum-Clarity LLCDate: April 10, 2026IBM Job ID: d7ckm0j0g7hs73dqvs10 (publicly verifiable)Backend: ibm_kingston (IBM Heron r3, 156 qubits, heavy-hex topology)Predecessor records: DOI 10.5281/zenodo.18498540 (February 5, 2026) · DOI 10.5281/zenodo.19478241 (April 9, 2026)

Plain Language Summary

What we found:

We ran an experiment to measure how long quantum information can survive inside the Y⊗Z orthogonal stabilizer subspace — the "protected plane" — before environmental decoherence degrades it below a useful threshold. Rather than a single decay curve, we used a three-curve discrimination design: a protected hold curve, an unprotected control, and a reset-and-reprepare control. We tracked four observables per module at each delay point across 13 delay times from 0 to 150 microseconds, on two modules selected from the April 9 Kingston run — one high-performing and one borderline.

The protected plane maintains fidelity above 80% until approximately 15 microseconds on both tested modules. Compared to an unprotected control state with equivalent gate budget, the protection advantage averages +61.9 percentage points on the good module — confirming categorically that the Y⊗Z orthogonal geometry provides coherence protection, not just a measurement artefact.

A non-monotonic revival feature appears in the protected curve at approximately 25–30 microseconds on both modules, consistent with the ~30 µs non-Markovian environmental memory timescale documented in the February 2026 QuantaCore reliability campaign. This structure is inconsistent with a simple monotonic exponential decay model.

Most operationally: the reset-and-reprepare curve consistently outperforms the passive hold curve across much of the tested delay window, with the largest advantage reaching +16.3 percentage points at 20 µs on the good module.

Why this matters:

This experiment closes the loop on three prior QuantaCore records. The February 5 record demonstrated topology-dependent error correlations. The February 6 record established that those correlations are non-Markovian with a ~30 µs characteristic timescale. The April 9 Kingston record validated the basis migration method at 116-qubit scale. This record now directly measures the coherence lifetime of the protected plane in the time domain, supports the non-Markovian revival interpretation suggested by the February campaign, and discovers a new operational strategy — periodic stabilizer refresh — that outperforms passive hold on present hardware.

Abstract

We report a Ramsey-style hold-time characterisation of the Y⊗Z orthogonal stabilizer subspace on IBM Kingston (ibm_kingston, Heron r3, heavy-hex, 156 qubits), using a three-curve discrimination design and four-observable tracking per module per delay point. Two four-qubit modules from the April 9, 2026 Kingston validation run were selected: a high-fidelity module (qubits 13, 14, 15, 19; F=91.6% at τ=0) and a borderline module (qubits 89, 90, 91, 98; F=86.1% at τ=0). Delay times of 0, 2, 5, 10, 15, 20, 25, 30, 40, 50, 75, 100, and 150 µs were applied, with dense sampling around the ~30 µs non-Markovian memory timescale identified in prior QuantaCore campaigns.

Three experimental curves were measured simultaneously: Curve A (protected hold — basis migration applied, delay inserted, observables measured), Curve B (unprotected control — comparable gate budget, no Y⊗Z stabilizer structure, delay inserted), and Curve C (reset-and-reprepare — basis migration applied, delay inserted, active qubit reset, basis migration re-applied, observables measured). Four observables were tracked per module per delay: primary |YZ_A(τ)|, secondary |YZ_B(τ)|, z_ortho(τ), and derived fidelity F(τ).

Key results:

The protected plane maintains F > 80% until approximately τ = 15 µs on both modules, with the good module showing a non-monotonic revival feature at τ = 25 µs (F rises from 70.2% at 20 µs to 76.8% at 25 µs before declining again). The fragile module shows a corresponding revival at τ = 30 µs (F rises from 56.2% at 25 µs to 67.6% at 30 µs). These revival features are inconsistent with simple Markovian exponential decay and are consistent with the ~30 µs non-Markovian environmental coupling timescale documented in the February 6, 2026 QuantaCore campaign. Stretched-exponential fitting yields β ≈ 1.3 for both protected curves, suggesting deviation from a simple passive dephasing picture and consistent with structured environmental coupling.

The average A-B protection advantage is +61.9 pp (good module) and +53.5 pp (fragile module) across all delay points, confirming that the Y⊗Z orthogonal subspace provides substantial and persistent coherence protection relative to an unprotected comparison state of equivalent circuit depth.

Curve C (reset-and-reprepare) outperforms Curve A (passive hold) across much of the tested delay window: on the good module, Curve C exceeds Curve A at τ = 2, 5, 10, 15, 20, 25, 30, and 50 µs, with the largest advantage of +16.3 pp at τ = 20 µs. The most natural interpretation is that passive hold allows correlated environmental error to accumulate continuously, whereas reset-and-reprepare periodically re-injects the protected state from a cleaner baseline. This does not eliminate environmental memory altogether — z_ortho drifts increasingly negative on Curve C at delays beyond 30 µs, reaching −0.36 at τ = 75 µs, indicating residual environmental influence accumulating across successive reset cycles — but it reduces its practical impact across much of the experimentally relevant window. The topology-dependence of the refresh advantage (stronger on the good module, smaller on the fragile module) connects to the February finding that error correlation structure is spatially localised.

Observable decomposition on Curve C reveals a dissociation pattern at long delays: secondary |YZ_B| remains stable at 61–67% from τ = 2 µs through τ = 150 µs while primary |YZ_A| decays from 98% to 9.5%, with z_ortho increasingly negative. This channel-specific failure signature — B stable, A decaying, z_ortho negative — has not been previously reported and is distinct from the failure modes observed in the static Kingston validation runs.

Non-Markovian environmental memory is confirmed on the good module: Curve C falls below Curve A at τ = 100 and 150 µs despite the reset operation, indicating that environmental influence is not fully erased by qubit reset and persists into subsequent preparation cycles on longer timescales.

This record establishes the fourth stage of a continuous QuantaCore research program: method disclosure (February 5, 2026), non-Markovian physical diagnosis (February 6, 2026), backend-adapted operational validation at scale (April 9, 2026), and direct protected-plane lifetime characterisation with mechanism discrimination (April 10, 2026, this record).

Keywords

Y⊗Z stabilizers, protected plane coherence, hold-time experiment, non-Markovian dynamics, revival feature, three-curve discrimination, orthogonal subspace, stabilizer refresh, IBM Kingston, Heron r3, quantum error characterisation, QuantaCore, basis migration, superconducting qubits, coherence lifetime, environmental memory, topology-dependent decoherence

1. Experimental Background and Continuity

This record is the fourth in a sequence of connected QuantaCore experimental disclosures:

The first record (DOI: 10.5281/zenodo.18498540, February 5, 2026) established the Y⊗Z parity-triangle consistency test and detected topology-dependent deviations from independence assumptions at 4.86σ and 3.76σ significance on IBM Heron r2.

The second record (February 6, 2026) extended that finding into a full non-Markovian reliability campaign reporting spatial correlations at 4.86σ, temporal memory at 2.8σ, and environmental persistence at 3.6σ, with a characteristic timescale of approximately 30 µs.

The third record (DOI: 10.5281/zenodo.19478241, April 9, 2026) validated the basis migration method at 116-qubit scale on IBM Kingston across 28 modules, with 89.39% average fidelity and 100% Z-orthogonality gate pass rate.

The present record directly measures the coherence lifetime of the Y⊗Z protected plane as a function of hold time, using a three-curve design and four-observable tracking to discriminate between failure modes and verify the non-Markovian revival prediction from the February campaign.

2. Experimental Design

2.1 Three-Curve Discrimination

Curve A — Protected hold: The Y⊗Z basis migration procedure is applied to the module, a variable delay τ is inserted on all four qubits simultaneously, and the stabilizer observables are measured. This is the primary measurement of protected-plane coherence as a function of hold time.

Curve B — Unprotected control: A comparison circuit with equivalent gate budget but no Y⊗Z stabilizer structure is applied, the same delay is inserted, and the same observables are measured. Any excess coherence in Curve A relative to Curve B is attributable to the orthogonal subspace protection specifically, not to general circuit properties or measurement configuration.

Curve C — Reset-and-reprepare control: The basis migration procedure is applied, the delay is inserted, all qubits are actively reset to the ground state, the basis migration procedure is re-applied, and the observables are measured. This curve isolates true hold-time decay from persistent environmental correlations: if Curve C matches Curve A, the decay is genuine hold-time decoherence; if Curve C is worse, environmental memory survived the reset operation.

2.2 Four Observables

Per module per delay per curve:

Observable
Role

Primary |YZ_A(τ)|
In-plane A-channel coherence

Secondary |YZ_B(τ)|
In-plane B-channel coherence

z_ortho(τ)
Leakage detector — rise indicates exit from protected plane

F(τ)
Derived fidelity = (z_success + yz_strength) / 2

2.3 Module Selection

Module
Qubits
τ=0 fidelity
Selection rationale

Good
[13, 14, 15, 19]
91.6%
Highest-fidelity module, April 9 run, z_ortho=0.000

Fragile
[89, 90, 91, 98]
86.1%
Borderline module, April 9 run, z_ortho=0.079

2.4 Delay Sweep

0, 2, 5, 10, 15, 20, 25, 30, 40, 50, 75, 100, 150 µs — dense sampling around the ~30 µs non-Markovian memory timescale.

2.5 Enhancement Stack

Identical to the April 9 enhanced run: resilience level 1 (TREX), Pauli gate and measure twirling (32 × 32 shots), XY4 dynamical decoupling, optimisation level 1 transpilation.

3. Key Findings

3.1 Protection Advantage

Average A-B gap of +61.9 pp (good) and +53.5 pp (fragile) across all delay points. The unprotected control never exceeds 30% fidelity. The protection advantage remains measurable throughout the delay sweep, even after the high-fidelity hold window has closed. The protection advantage is persistent, not transient.

3.2 Non-Markovian Revival Confirmed

Both modules show non-monotonic revival features in Curve A centred on 25–30 µs:

Good module: 70.2% at 20 µs → 76.8% at 25 µs → 65.4% at 30 µs

Fragile module: 56.2% at 25 µs → 67.6% at 30 µs → 51.5% at 40 µs

This structure is inconsistent with a simple monotonic exponential decay model. Stretched-exponential fitting yields β ≈ 1.3, consistent with accumulating environmental correlations. The revival timescale matches the ~30 µs characteristic memory timescale documented in the February 6 campaign, extending that finding directly into the time domain within the Y⊗Z basis.

3.3 Stabilizer Refresh Strategy

Curve C outperforms Curve A across much of the tested delay window. The largest advantage is +16.3 pp at τ = 20 µs on the good module. The result suggests a new operational strategy: periodic stabilizer refresh cycles may outperform passive hold on present superconducting hardware. This is loosely analogous to cache refresh — implemented here as repeated protected-state re-injection rather than passive retention.

The safe claim from the current data is: for much of the 2–50 µs window, reset-and-reprepare outperforms passive hold, especially on the higher-quality module. This is not yet a demonstration that arbitrary stored information can be preserved through refresh cycles — that would require an additional experiment encoding specific logical content into the protected plane and verifying its recovery after refresh.

The topology-dependence of the refresh advantage (stronger on the good module, attenuated on the fragile module) suggests that local hardware calibration quality determines how much benefit periodic reprepare provides — consistent with the spatial error correlation structure documented in the February campaign.

3.4 Observable Dissociation Signature on Curve C

At long delays, Curve C shows a distinctive channel dissociation: |YZ_B| stable at 61–67%, |YZ_A| decaying to near zero, z_ortho increasingly negative (−0.36 at τ = 75 µs). This specific failure signature — B-channel stable, A-channel collapsed, z_ortho negative — appears distinct from the failure modes observed in prior Kingston validation runs and may reflect asymmetric accumulation of environmental phase across successive reset-and-reprepare cycles. This is reported here as a new finding warranting further investigation.

4. IP Statement

The Y⊗Z basis migration procedure implemented in this experiment is protected under U.S. Patent Application No. 19/643,807 (filed April 10, 2026), which claims benefit of Provisional Application No. 63/952,786 (filed January 2, 2026). This dataset constitutes experimental characterisation of the protected-plane coherence lifetime of that method on IBM Kingston hardware. Implementation scripts are proprietary and are not included in this dataset.

5. Reproducibility and Verification

IBM Quantum job ID: d7ckm0j0g7hs73dqvs10 — all results independently verifiable.

Files in this record:

yz_holdtime_20260410_121317.json — complete per-module, per-delay, per-curve results including raw expectation values, derived observables, fitting parameters, and three-curve discrimination metrics

yz_holdtime_20260410_121317.png — six-panel visualisation: three-curve F(τ) comparison, four-observable breakdown, and A-B protection advantage, for both modules

6. Acknowledgements

IBM Quantum for hardware access and QPU credits on ibm_kingston.

© 2026 Amit Brahmbhatt, Quantum-Clarity LLC. Data: CC BY 4.0.