Sample audit report
This is a real T5.4 verdict against a real public claim, shown in full — including the rules that did not pass. An audit that only ever returns good news is a marketing page, not an audit.
eig-t54-ionq-879bconditional_pending_reexecutionBell-state fidelity of 87.9(7)% after frequency conversion, at an entangled pair rate of 1,032 ± 1 pairs/s.
“Quantum Networking at the Speed of Quantum Computation” — preprint, 7 Oct 2026, 39 authors (IonQ + Duke). arXiv:2610.10705
N2 — vendor-calibrated simulation with canonical hash and full verdict.
2 pass · 2 partial · 1 fail · 4 open — of 9 truth rules.
Verdict rationale
The claim is arithmetically sound as stated. The reported uncertainty is recoverable from the claim itself, the measurement is scoped unambiguously, and the headline number is not overstated. Nothing here suggests the authors misrepresented their result.
What the claim does not yet establish is generalisability. It was measured on one platform, against a classical baseline that was never stated, with no independent re-execution, and without a canonical hash over a runnable artifact. Six of nine rules are therefore PASS, PARTIAL or OPEN rather than conclusive.
Separately, and importantly: a 12.1% error rate is roughly two orders of magnitude above the ~0.1% error rate fault-tolerant designs assume. The 87.9% figure is a real and useful result. It is not fault-tolerant performance, and an audit that let a reader infer otherwise would be worthless.
The 9 truth rules, scored
T5.4 scores a claim against a fixed set of nine rules. The rules do not change between claims, and they are not weighted — a claim passes or it does not, rule by rule, and the reader sees every one.
TR-1PASSBaseline definition is unambiguous. State the trivial comparator.
Bell-state fidelity is named explicitly, and the measurement is scoped to the post-frequency-conversion state. A reader can identify what was measured without inferring it from context.
TR-2PARTIALClassical solver comparison present (when algorithm has a classical analog).
Entangled pair rate has a direct classical analog (correlated photon-pair sources). The claim is not framed against the classical rate, so the comparison is absent rather than negative. A reviewer cannot tell whether the gap was considered.
TR-3OPENVendor-specific hardware execution on 3+ vendors, or documented why fewer.
Executed on one platform (IonQ) with a Duke collaboration. Fewer than three, and the artifact does not document why. This is the single largest gap in the claim.
TR-4PASSStatistical significance ≥ 3σ. Sample size and seed disclosed.
The fidelity is reported as 87.9(7)% — the parenthetical uncertainty makes the confidence interval recoverable from the claim itself, and the rate is reported as 1,032 ± 1 pairs/s. A reader can reconstruct the error bar without contacting the authors.
TR-5OPENCalibration timestamp within vendor-stated stability interval.
Not determinable from a published artifact. This rule is satisfiable only by a re-execution against a dated calibration record, which is what lifts an audit from N2 to N3.
TR-6OPENCross-vendor portability demonstrated.
Not demonstrated. Portability is exactly what the vendor-neutral orchestration layer is designed to test, and it is the reason this claim was selected as a first audit rather than dismissed.
TR-7OPENIndependent re-execution by a second engineer.
Not yet performed. This is a scheduling constraint, not a property of the claim. It is the rule that converts a paper audit into a hardware audit.
TR-8FAILAdversarial test passed (noise injection, gate failure).
The published result sits at 12.1% error against the roughly 0.1% error rate that fault-tolerant designs assume — two orders of magnitude off. The claim itself is not wrong, but it is not fault-tolerant performance and no adversarial test in the artifact establishes otherwise.
TR-9PARTIALCitation form is citable (canonical hash + reproducible source).
The source is a preprint with 39 named authors and a fixed date, so it is citable in the ordinary sense. It is not reproducible in the T5.4 sense: there is no canonical hash of a runnable artifact, which is precisely what an EQL manifest supplies.
How far this evidence can be pushed
An audit is only as strong as the evidence behind it. The N-tiers make the distinction explicit rather than leaving it to the reader to infer.
N0Claim only, no runnable artifact. Not audit-able. Reject.
Not reachedN1Simulator only, canonical hash, basic verdict.
ReachedN2Vendor-calibrated simulation, canonical hash, full verdict.
ReachedN3Real hardware on 1 vendor, canonical hash, full verdict, cross-portability.
Not reachedN4Real hardware on 3+ vendors, canonical hash, full verdict, independent re-execution.
Not reachedWhat would lift this verdict
An OPEN verdict is a work order, not a dead end. Each remaining gate names the specific action and the specific rule it closes.
Reconstruct the experiment in EQL and emit a canonical hash over the runnable artifact.
closes TR-9 · PARTIAL → PASS
Re-execute on one physical vendor backend against a dated calibration record.
closes TR-3, TR-5 · N2 → N3
Re-execute on a second, architecturally distinct vendor backend.
closes TR-3, TR-6 · portability demonstrated
Independent re-execution by a second engineer against the same canonical hash.
closes TR-7 · N3 → N4
Adversarial pass: inject noise and gate failure, report where the claim stops holding.
closes TR-8 · FAIL → resolved
Why this matters commercially
Every one of the open gates above is a hardware access problem. That is the entire commercial argument for a vendor-neutral orchestration layer: T5.4 needs to run the same canonical artifact on more than one backend, and no vendor will ever provide that for you. The layer that does is the layer that gets paid.
Submit your own claim
One claim, one verdict, one signed audit ID. The same nine rules, whether the claim comes from a vendor, a university, or your own team.