Fixed-order postselected CHSH reference acquisition for parity-constrained spatial-mode qubits on a commercial cloud photonic processor

Martin—quick take on arXiv:2608.18153. The useful content is real: eight sequential postselected CHSH passes on Quandela's Belenos, all raw S>2, with archived counts, job identifiers, submitted-target hashes, and verification code, plus a careful separation of conditional shot noise, between-pass dispersion, and modeled systematics. The paper earns credit for reporting raw and reweighted values side by side, for printing the marginal diagnostics that go against the clean model, and for explicitly scoping the result as an operational reference acquisition rather than an entanglement witness or loophole-free Bell test. The Appendix A reduction to the single-mode RLBW CZ is clean and machine-checked at the submitted-target level.

The soft spot is exactly where the stress-test note lands. The raw numbers are what they are: on returned, postselected counts, S=2.485±0.019 combined. What is not established is that those numbers are CHSH scores of the intended two-qubit circuit. The provider never returns the compiled physical mapping, and the paper's own acceptance-fraction comparison disagrees with the submitted-target model: measured fractions track acquisition position, not the predicted setting pattern. Since setting order was fixed, position and setting are perfectly confounded, and Eq. (16) shows a linear drift adds +2β with no model-independent correction. The reweighted weakest pass, 2.040±0.062, overlaps 2 within 1σ. None of this is hidden—the limitations section and Table VIII say so—but it means the 'all eight raw passes above 2' headline is not the same as eight independent confirmations of CHSH in the intended circuit.

The reweighting scene is also softer than it looks: kappa values are session-start dashboard metadata, not contemporaneous calibrations, and the marginal-consistent value is chosen to zero a residual. The authors label this an 'archived-metadata model scenario,' so it is not fraudulent; it just should not be quoted as a corrected score. The self-citation to the companion paper for the encoding vocabulary is unproblematic; the CHSH data here are distinct and openly archived.

Bottom line: this is a serious, reproducible reference acquisition with unusually honest limitations. A thoughtful referee should engage. I would send it to peer review with the request that the abstract and headline either define S strictly on returned counts or carry the compiler caveat prominently; the body already does. For anyone working on cloud photonic benchmarking or auditable user-level protocols, it is worth a careful read. It is not a Bell experiment and should not be positioned as one.

The eight raw per-pass values and Scount are internally consistent and reproducible from the archived counts, so the narrow claim that these numbers were observed holds. What does not hold without further evidence is the identification of those numbers as CHSH scores of the intended encoded two-qubit circuit. Eq. (3)'s factorization is verified only for the submitted 16x16 matrices (Table X, Appendix A); the server-side compiled transformation is never returned. The paper's own in-corpus check of the executed process, the accepted-fraction comparison in Sec. V, disagrees with the submitted-target model: measured fractions (0.147, 0.175, 0.169, 0.179) do not follow the predicted b0/b1 pattern (0.161/0.141 at fitted v about 0.94) and instead track acquisition position. Because setting order is fixed, position and setting are perfectly confounded, and Eq. (16) shows that a linear drift contributes plus 2 beta to S with no model-independent correction. The archived reweighting already drops the weakest pass to 2.040 plus or minus 0.062, overlapping 2 at 1 sigma. Thus a positive drift or setting-dependent compilation could be why all eight raw S exceed 2, and the archived data cannot exclude that. The paper's scoping language mostly acknowledges this, so the concern is about the load-bearing interpretation of 'raw postselected CHSH score,' not about arithmetic honesty.

The paper is careful and unusually honest about its limitations: it explicitly disclaims entanglement-witness, loophole-free, and cross-platform claims, reports the between-pass dispersion through Q/nu, separates count-pooled descriptors from pass-level analyses, and archives counts, hashes, and code. The raw S values are internally consistent: the count-pooled 2.485 plus or minus 0.019 follows from the four tabulated correlators, the session means and sample standard deviations match the per-pass table, and the archived package is structured so an independent user can recompute the tables. The central claim as scoped, 'we observed these raw operational scores in this fixed-order same-day sequence,' is therefore well supported. The load-bearing weakness is the step from 'counts on user-labeled circuit modes' to 'CHSH score of the submitted encoded two-qubit circuit.' That step requires the provider's compiled transformation to preserve the user-mode labeling and the block-local structure of Eq. (3), and the compiler never returns the executed mapping. The paper itself flags this repeatedly, and the reader's conditional verdict already rests on it. I add one aggravating observation: the only in-corpus quantitative check of the executed process, the accepted-fraction comparison, does not confirm the submitted-target model; the measured fractions are inconsistent with the predicted pattern and track acquisition position. Combined with the perfect setting-position confound and the drift term in Eq. (16), this means the margin above 2 in the raw series is not robust to plausible order-dependent or compilation-dependent effects. However, because the authors present the result as an operational reference acquisition rather than a claim about an underlying two-qubit state, this concern does not overturn the central claim; it strengthens the conditionality already expressed in the verdict. I therefore leave the verdict unchanged and agree with the reader's identification of the compiled-mapping assumption as the weakest point.

If the narrow claim is accepted, the paper's value is methodological: it provides a fully archived, reproducible acquisition-and-reporting protocol for third-party characterization of a cloud photonic processor, with explicit separation of count-pooled descriptors, pass-level dispersion, and model-scenario reweighting. Strengths include machine-checked verification scripts, content hashes for submitted targets, a clean analytic reduction in Appendix A with numerical verification to machine precision, and unusually candid reporting of excess dispersion, residual marginals, and the unverified compiled mapping. The result does not establish an entanglement witness, and the authors say so; its significance lies in the auditable reference-acquisition framework rather than in a new Bell-type violation.

The paper aims to establish that a postselected CHSH score above the classical bound of 2 can be obtained by external users of a commercial cloud photonic processor, using photonic qubits encoded in parity-constrained spatial-mode subspaces. It reports eight sequential four-setting passes, each with raw $S>2$, with session means 2.40 and 2.58 and a count-pooled $S_{\mathrm{count}}=2.485\pm0.019$ from 26,833 accepted coincidences. The authors deliberately scope the result as an auditable reference acquisition: the compiled physical mapping is not returned, setting order was fixed, and residual remote-setting marginals remain, so the data do not support an entanglement-witness or cross-platform benchmarking claim. If the protocol holds up, it gives other external users a reproducible template for characterizing black-box photonic hardware with explicit efficiency and provenance diagnostics.

What carries the argument. The load-bearing object is the submitted target unitary factorization $U^{\mathrm{target}}_{xy}=\bigl(U^A_{\mathrm{read}}(x)\oplus U^B_{\mathrm{read}}(y)\bigr)\,U_{CZ}\,\bigl(U^A_{\mathrm{prep}}\oplus U^B_{\mathrm{prep}}\bigr)$ of Eq. (3): two eight-mode single-photon blocks in the zero-sum subspace $N=\{v\in\mathbb{C}^8:\sum_i v_i=0\}$, with logical rails $|0\rangle_L=(e_0-e_1)/\sqrt{2}$ and $|1\rangle_L=(e_2-e_3)/\sqrt{2}$, coupled by a postselected linear-optical controlled-$Z$ of the Ralph-Langford-Bell-White type with ideal success probability $1/9$. The gate works because pairwise rail couplings with identical transmissivity $T=1/3$ preserve the antisymmetric rail combination and reproduce the single-mode controlled-$Z$ amplitudes exactly, giving a postselected logical transfer matrix $\tfrac13\,\mathrm{diag}(1,1,1,-1)$ and ideal $S=2\sqrt{2}$ at the chosen settings. The analysis then assembles $S=E(a_0,b_0)+E(a_0,b_1)+E(a_1,b_0)-E(a_1,b_1)$ from accepted four-way coincidence counts, with each four-setting pass as the experimental unit.

Core claim. The central claim is that the operational CHSH score $S$ computed on accepted logical coincidences exceeded 2 in all eight passes of one same-day, two-session sequence on a commercial 24-mode photonic processor, giving session means 2.40 and 2.58 and a count-pooled $S_{\mathrm{count}}=2.485\pm0.019$. The authors state that the score is well-defined on returned counts under any physical mapping, but interpreting it as the CHSH score of the intended encoded two-qubit circuit requires the unverified assumption that the server-side compiled transformation preserves the submitted-target factorization of Eq. (3). Efficiency reweighting by archived port-transmittance ratios moves the pooled value to $S^{\mathrm{rw}}_{\mathrm{count}}=2.380\pm0.021$, with the weakest reweighted pass at $2.040\pm0.062$, overlapping 2 within $1\sigma$. The paper's explicit conclusion is that this is a fixed-order postselected reference acquisition, not an entanglement witness, loophole-free Bell test, or cross-platform benchmark.

The premise it rests on. The load-bearing premise is that the cloud processor's internal implementation of each submitted circuit keeps the intended block-local structure—the same preparation on both qubits, separate readout rotations, no coupling between the two blocks' halves, and the same rule for keeping events—and if it does not, the reported score is still well-defined on the returned counts but does not measure the intended CHSH target.

If the result stands, external cloud users can run a complete, auditable CHSH reference acquisition with archived job identifiers and content-hashed target matrices, rather than relying only on vendor-reported processor metrics.

The same-day sequence produced raw $S>2$ in every pass, but the excess between-pass dispersion ($Q/\nu=6.1$) means a single count-pooled shot-noise error is not a pass-level precision; future reports should keep the pass as the unit of analysis.

The efficiency-reweighted model scenario, $S^{\mathrm{rw}}_{\mathrm{count}}=2.380\pm0.021$, brackets the dominant identified readout imbalance, but its weakest pass overlaps 2, so the raw and reweighted series must be reported together.

Because setting order was fixed and the compiled mapping was not returned, the dataset cannot validate a common-state, local-analyzer interpretation; it serves as a like-for-like reference for repetitions using the same target and acquisition order.

If the service ever exposed compiled unitaries, the same protocol would likely convert a reference acquisition into an actual verification of the executed circuit; that step is absent here.

The persistence of the block-A marginal at setting $a_1$ after efficiency reweighting suggests a setting-dependent effect such as compilation differences, drift, or cross-block mixing, and a swapped-label control would be the decisive test.

The same parity subspace could be upgraded from postselection to genuine error detection by measuring the zero-sum syndrome, which the paper explicitly does not do.

A counterbalanced setting order would not only remove the identified confound; it would turn the drift term $2\beta$ of Eq. (15) into an estimable parameter rather than a limiting uncertainty.

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review