Phase B reproducibility package: count records, job manifest, and analysis code for the fixed-order postselected CHSH reference acquisition on the Quandela Belenos processor

Supplemental material for "Fixed-order postselected CHSH referenceacquisition for parity-constrained spatial-mode qubits on a commercialcloud photonic processor" (Tully, Washburn, Simons, 2026). This package contains the complete archived cloud-job records, job-IDmanifest, circuit-construction and analysis code, content hashes for thefour submitted target unitaries, and verification scripts that recomputeevery table and numerical claim in the manuscript. Every number can beregenerated from the archived records without live cloud access. Results are framed as a fixed-order, single-provider CHSH-score referenceacquisition plus reporting protocol, under submitted-target factorizationwith unaudited server-side compilation and a fair-sampling assumption. Thelead quantity is the operational CHSH score S on accepted logicalcoincidences. The primary analysis treats each complete four-setting passas the experimental unit: the per-pass raw scores have session means 2.40and 2.58 with sample standard deviations 0.15 and 0.03, and the eightpasses do not share a common mean (chi^2/nu = 6.1), so no count-pooledshot-noise error is a platform precision. The count-pooled raw andfixed-ratio reweighted scores are therefore reported as model scenariosonly. The efficiency reweighting is an archived-metadata model scenario,not a corrected platform score: it uses start-of-session transmittancemetadata whose temporal stability was not re-measured, and theaccompanying kappa ranges are reported as an ad hoc stress scan ratherthan a calibrated uncertainty band (narrow +/-10% metadata band2.356-2.406; wider mis-specification scan 2.341-2.441 for kappa in[1.2, 1.8]). Every returned pattern has already passed the provider's uncharacterizedmin_detected_photons returned-sample filter, so all scores are conditionalon that prior server-side selection stage, which could not be auditedthrough the interface; raw source-clock shot counts are not exposed, so noquantity can be re-expressed on a per-trial denominator. Because the data are one dependent same-day sequence on a single provider,the deposit supports a reporting protocol plus a reference acquisition, nota validated cross-provider comparative metric. Standard CHSH-to-entanglementconversions (concurrence, entanglement of formation, and self-testing /extractable-fidelity bounds) are not reported, because the residualremote-setting marginals do not establish their common-state,local-analyzer premises. Values regenerated by this package- Per-pass raw S for all eight passes; session means 2.40 (sample SD 0.15) and 2.58 (sample SD 0.03); Cochran heterogeneity chi^2/nu = 6.1 on the raw pass estimates- S_count = 2.485 +/- 0.019 (conditional shot noise) on 26,833 accepted coincidences -- a conditional descriptor, not a platform precision- S^rw_count = 2.380 +/- 0.021 (conditional), session-specific block-A reweight with kappa_A,1 = 1.524 and kappa_A,2 = 1.587; weakest reweighted pass 2.040 +/- 0.062, which overlaps S = 2 within 1 sigma- Equal-session S(kappa) scan: S decreases monotonically from 2.485 at kappa = 1 to 2.315 at kappa = 2, reaches S = 2 only near the implausible kappa ~ 12, and tends to 1.89 as kappa -> infinity- S^rw_count kappa narrow metadata band 2.356-2.406 (+/-10% about archived ratios); wider mis-specification stress scan 2.341-2.441 for equal-session kappa in [1.2, 1.8]; marginal-consistent kappa_eff = 1.395 gives 2.404- S^rw_AB = 2.376 +/- 0.021 when both blocks are reweighted- Residual remote-setting marginals: raw b1 +0.100 +/- 0.008, a1 +0.037 +/- 0.008; after reweighting b1 -0.033 +/- 0.009, a1 +0.038 +/- 0.009. The a1 residual is essentially unchanged by the reweight and is therefore irreducible under any single fixed ratio.- Submitted-target checks: 16x16 unitarity residuals <= 7.8e-16 and postselected logical 4x4 transfer matching (1/3)diag(1,1,1,-1) to 1.7e-16; postselection-predicate digest prefix 6bae83187a6bba37 (complete classify source plus PORT_A/PORT_B). These concern the SUBMITTED targets only; the server-side compiled physical mapping is never returned and cannot be verified. Contents- data/raw/: full click-pattern count records and platform calibration metadata for the two CHSH acquisition sessions of 4 July 2026 (32 CHSH jobs across eight passes, plus a two-job photon-recycling A/B comparison) and the exploratory gate-characterization session- manifest/job_manifest.csv: all 59 cloud job IDs mapped to the manuscript table or figure each supports, with setting order and timestamps. The manifest is reproduced in full in the manuscript appendix, so every reported quantity is traceable to an identified job.- scripts/verify_targets.py: the single pass/fail verification entry point. Recomputes the SHA-256 hashes of the four submitted 16x16 CHSH target unitaries (under the documented 12-decimal canonical serialization) and of the complete offline postselection predicate; checks the 16x16 unitarity residuals and the postselected logical 4x4 transfer against (1/3)diag(1,1,1,-1); regenerates S_count and S^rw_count from the archived counts; compares the measured per-setting accepted logical fractions against predictions from the archived targets at v = 1, the fitted v, and the fully distinguishable limit; and asserts that no numerical table is imported from the companion single-block study. Exits non-zero on any mismatch.- scripts/: circuit construction and postselection/correlator analysis; the hardware submission driver used for the CHSH acquisition sessions; the count-pooled descriptor S_count; pass-level random-effects / Birge-Student sensitivity summaries (DerSimonian-Laird/REML between-pass variance, Hartung-Knapp interval -- implemented directly, with no external meta-analysis package, and kept as sensitivity analyses only, because the passes are sequential and not asserted independent); fixed-ratio readout-imbalance reweighting reported as S^rw_count (multinomial delta-method error, bootstrap-checked) with an exploratory reweighted pass-level summary S^rw_pass and the kappa stress scan; remote-setting marginal diagnostics; acquisition-position summaries that are descriptive only and are not interpreted as temporal trend tests (setting and acquisition position are perfectly confounded); the exploratory two-setting ZZ/XX witness with uncertainties (an archived cross-check, not a headline result); a two-component distinguishability model; and verify_package.py, which recomputes the manuscript tables from data/raw and checks them against the quoted values- requirements.txt: pinned software environment (tested with CPython 3.11.10, numpy 2.4.4, perceval-quandela 1.2.4)- SHA256SUMS.txt: SHA-256 checksums for package files Quick startpython3.11 -m venv .venv && source .venv/bin/activatepip install -r requirements.txtcd scriptspython3 verify_targets.py && python3 verify_package.py && python3 revision_stats.py Run verify_targets.py first: it is the single gate that returns exit code 0only if the target hashes, ideal-target algebra, headline scores, andcompanion-isolation assertions all match the manuscript. File integrity: shasum -a 256 -c SHA256SUMS.txt Verification itself performs no network calls. The advertised verificationpath requires both pinned dependencies (numpy and perceval-quandela):scripts import the circuit-construction module, which uses perceval forphotonic BasicState bookkeeping. ProvenanceAll jobs were submitted to qpu:belenos through the public Perceval cloudinterface as standard external-user jobs under Quandela's publicly availableExplorer offer. Job IDs in the manifest can be cross-checked against theQuandela cloud job console. No credentials, tokens, or live URLs areincluded. LicenseCode/data split: source code under scripts/ and this documentation arereleased under the MIT license (LICENSE); archived data records under data/are released under Creative Commons Attribution 4.0 International, CC BY 4.0(LICENSE-DATA). If you use either, please cite the accompanying manuscriptas the preferred attribution.