Publication validation scope

This pre-release supports executable workflows at different maturity levels. Implemented APIs, passing correctness checks, historical measurements, and research goals are distinct. The capability catalog and known limitations remain authoritative for each path; this review does not promote their maturity levels.

Evidence reviewed on 2026-09-11

Area	Implemented and checked	Boundary
Local learning	Circuit construction, PyTorch Module/autograd, optimizers and local measurements; Python 3.10, 3.11 and 3.12 core CI passes	Gradient support depends on representation and backend; no blanket higher-order-gradient claim
GPU and distributed execution	A800 single-GPU gate checks; two-GPU statevector forward, reverse and training; MPS training on 2, 4 and 8 GPUs with checkpoint/resume	Selected correctness workloads, not a performance comparison or capacity certification
Cross-node statevector	Two nodes, one A800 per node: forward, gradients and training/resume over NCCL/TCP	The forward probe uses five qubits; no RDMA, throughput, strong-scaling or large-state capacity claim follows
Tensor networks	Native slicing and distributed CPU correctness checks	The A800 results above do not certify all tensor-network execution or gradient paths
Jiuding and Quafu	Experimental adapters, submission/result contracts and runnable credential-dependent examples	No new live QPU job was submitted in this review; earlier provider evidence remains limited to its recorded task and environment
Digital twin	Calibration-conditioned predictions and program/task-bound comparisons	Retrospective replay is not general prospective prediction across circuits, devices or calibration epochs
QEC and FTQC	Local three-data-qubit repetition-memory experiments with syndrome, decoder and correction records	General logical operations, threshold claims and real-time QPU feedback remain research goals

The reviewed CI run passed all 15 jobs. Its coverage job passed 2,455 tests, skipped 36 and measured 66.10% line coverage under the repository's existing .coveragerc exclusions. The CPU distributed tier passed 402 tests with 28 skipped; a separate benchmark-contract tier passed 113 with 12 skipped. These counts are different scopes and must not be added into a unique-test total. Skipped tests are not evidence of supported hardware or behavior.

The A800 environment was Python 3.12.13, PyTorch 2.13.0+cu130, CUDA 13.0 and NCCL 2.29.7, on A800-SXM4-80GB devices. Raw logs remain in the maintainer's private validation archive because they contain infrastructure metadata. This summary is not a publicly replayable performance artifact or a scalability release gate.

One CI checkpoint-lease fault test failed without captured worker diagnostics. The harness now exposes worker output and separates normal checkpoint deadlines from deliberate timeout injection. All 14 fault scenarios passed on Linux, and the final CI passed; the original failure's root cause was not established. An additional macOS distributed check hit launcher timeouts and was stopped; this review does not certify macOS distributed execution.

Reproduce the relevant checks

Use the dependency installations and tier commands in CI and the testing manual. Hardware scenarios live in distributed tests, including statevector_forward_executor.py, statevector_reverse_executor.py, statevector_training_runtime.py, and test_mps_training_runtime.py. The distributed MPS runbook provides workload and evidence interpretation guidance.

Historical benchmark figures retained in Git history are records of their original experiments. They do not certify the present release. Performance comparisons require matched workloads, precision, measurement methodology and auditable artifacts; FTQC and general predictive digital-twin claims require their own scientific evidence.