A preprint study of quantum-hardware workloads found that a deletion-recovery result on IQM Emerald did not repeat at a later execution snapshot. A separate test on IonQ Forte Enterprise 1 passed both of its frozen criteria in two task windows, making the findings a split result rather than a single verdict on quantum machines.
On IQM, the recovery return probability fell from 0.7384375 in discovery, based on 4,726 successes out of 6,400 shots, to 0.6240625 in replication, based on 3,994 of 6,400. The reported replication lower bound was 0.607086. The frozen confirmation rule required at least 4,376 successes, and the later run missed that mark by 382.
The recovery analysis was frozen locally before confirmation, but the paper says it was not a formal public preregistration. The decision is therefore a result for this specified workload and contract, not a general claim about the processor.
What was being tested
The study asked whether a permissive round-trip screen changes as route length grows, whether route-aligned task contrasts repeat at a later snapshot, and how recovery behaves under a separate virtual-wire contract.
RTSE is that round-trip screen: it prepares an input at a route root, swaps it to a remote endpoint and back, unprepares it at the root, and measures only the root output. Returned ancillas are deliberately not checked, so a high RTSE score does not by itself prove that the remote endpoint was reached or predict teleportation or recovery performance.
The recovery circuit crossed every tetrahedral input with four deletion positions. Deletion was implemented by parking and marginalizing the carrier - an ideal partial-trace wrapper - rather than by physically losing a qubit or ion.
The workloads ran on two processors: IQM Emerald used explicit physical-qubit and native PRX-CZ compilation, while IonQ Forte Enterprise 1 used five submitted virtual wires under an advertised all-to-all contract. Across eight tasks, all 624 program executions and 101,600 requested shots completed. Because IonQ physical-ion assignments were not returned, this was an architecture-stratified rather than controlled comparison.
IQM showed why aggregate scores can hide movement
In the IQM communication test, the discovery snapshot passed 27 of 28 route-performance endpoints and replication passed all 28, yet none of four confirmatory temporal targets met the replication rule.
Aggregate RTSE estimates differed by 0.00125, but the RTSE-teleportation penalty contracted from 0.04625 to 0.01422 as the teleportation estimate rose by 0.03328. On route 39 in direction 1, the penalty fell from 0.21125 to 0.04500. The close aggregate score therefore did not establish temporal equivalence or guarantee that the selected confirmatory contrasts would repeat.
A separate IQM length study found both RTSE and the predecessor's root-marginal measure declined as the route-length setting rose from L = 2 to L = 10. With equal-window weighting, the drops were 0.18625 for RTSE and 0.10594 for the predecessor; the protocol-by-length interaction was -0.08031, so the prespecified RTSE-retention advantage was not supported. In the two windows separately, RTSE drops were 0.17063 and 0.20188, versus 0.08062 and 0.13125 for the predecessor, with interactions of -0.09000 and -0.07063.
To quantify uncertainty in the communication comparisons, the analysis used conditional simultaneous 95% bounds across 100 fixed program cells, drawing 100,000 parametric binomial bootstrap replicates from smoothed cell proportions.
IonQ cleared a different set of tests
On IonQ, recovery means were 0.910625 in window A and 0.922500 in window B. The corresponding lower bounds were 0.876672 and 0.888547. Recovery-minus-adjoint contrasts were 0.446250 and 0.442500, with lower bounds of 0.398234 and 0.394484. Both windows passed the study's two frozen criteria.
Each recovery cell was paired with an adjoint-decoder control using the same input and deletion labels. That control was not an uncoded baseline and did not isolate gate order from the direction of rotation parameters, so the positive contrast is not a coding-gain result.
A workload diagnostic, not a ranking
Taken together, the authors frame the findings as execution-workload diagnostics indexed by workload, placement or virtual-wire contract, compilation, architecture, and execution snapshot. They do not present them as evidence of coding gain, error suppression, physical qubit or ion loss, fault tolerance, or an architecture ranking.
The study's deletion wrapper does not reproduce physical qubit or ion loss, and the differing compilation and mapping contracts make the IQM and IonQ results unsuitable for a direct performance league table. They describe what happened in these observed workload cells and execution snapshots.
The manuscript is a preprint identified as arXiv:2608.26010v1 [quant-ph], dated 26 Aug 2026. A secret-free reproducibility dataset is publicly available on Zenodo at doi:10.5281/zenodo.21969397, with submitted QASM, derived data, analysis and figure-generation code, figures, manifests, checksums and validators; raw provider records and credentials are excluded.
No funding source is reported. The acknowledgment says OpenAI Codex assisted with prose and analysis-related code, while the authors retain responsibility for the research questions, decisions, interpretation and manuscript.
Paper data and sources
Original title: From Round-Trip State Echo to Error Recovery: Snapshot-Resolved Quantum-Hardware Diagnostics
Authors: Isaac Barouch Essayag, Aryeh Lev Zabokritskiy
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text