This page consolidates the public technical material: package identity, module map, decoder contracts, mathematical invariant, performance trade-offs, deployment boundaries, commercial license limits, DOI provenance, and the planned Workbench artifact manifest.
QECTOR is a source-available platform, not a single decoder script. A Rust core is exposed through PyO3 and Maturin, then wrapped by a NumPy-friendly Python API with Stim/DEM, PyMatching, Sinter, LDPC, benchmark, and backend-selection support.
Every decoder is evaluated first against the same parity-check contract:
H · correction == syndrome (mod 2)Accuracy, speed, memory, and commercial claims come after this correctness gate and must be tied to a reproducible artifact.
| Subsystem | Role | Public boundary |
|---|---|---|
| Union-Find | Synchronous odd-cluster growth, union-by-rank fusion, CSR-style adjacency, and spanning-forest peeling. | Fast syndrome-faithful decoding on supported matching-graph-style codes. Not an optimality claim. |
| Exact Blossom MWPM | Builds augmented defect graphs, solves sparse MWPM, then replays qubit paths. | LER parity with PyMatching on checked larger-distance Stim workloads. PyMatching remains the latency leader on common exact-MWPM workloads. |
| Sparse Blossom | Region-growing restricted matching with exact fallback when validity fails. | Near-optimal and syndrome-faithful on tested cases. Not advertised as exact MWPM. |
| BP-OSD | Min-sum belief propagation followed by reliability-ordered GF(2) solving. | Useful for LDPC/qLDPC workflows. Claims must specify code family, noise model, seed, and artifact. |
| Belief-Matching | Uses detector-error-model hyperedge information to update matching weights. | Accuracy mode for selected correlated workloads. Slower than plain MWPM. No universal superiority claim. |
| Batch workflows | Batch execution and backend validation paths. | Throughput claims require local artifacts with hardware, package version, circuit, seed, and hashes. |
QECTOR weighted Blossom matches PyMatching logical error rate on checked rotated-surface-code Stim workloads through d=15, while remaining much slower per shot.
On selected correlated workloads, belief-matching reports lower observed LER at selected distances, with runtime overhead.
Public claims should point to GitHub artifacts, benchmark_results, package version, commit SHA, and the Zenodo DOI.
The Workbench roadmap should normalize circuits, detector models, code geometry, decoder configuration, empirical metrics, environment data, commit SHA, package version, DOI, and SHA-256 hashes into one exportable report bundle.
{
"repository": "https://github.com/GuillaumeLessard/qector-decoder",
"package": "qector-decoder-v3",
"version": "0.5.1",
"doi": "https://doi.org/10.5281/zenodo.20825980",
"backend": "union_find | blossom | sparse_blossom | belief_matching | bp_osd | batch",
"shots": 3000,
"observed_logical_error_rate": 0.005,
"artifact_hashes": {
"results_json": "64 hex characters",
"results_csv": "64 hex characters"
}
}