Benchmark Evidence

Reproducible QEC decoder validation

QECTOR benchmarks are presented with scoped claims: tested circuits, distances, seeds, error models, latencies, memory profiles, and artifact hashes. The strongest public result is not a fastest-decoder claim; it is reproducible evidence with PyMatching LER parity, belief-matching accuracy mode, BP-OSD support, and CPU/GPU consistency checks.

Snapshot

Latest validated summary

832tests collected

829tests passed

d=15LER parity audit vs PyMatching

GPUCUDA/OpenCL bit-identical on tested batches

Report build: git 729282f. License: QECTOR Decoder Source-Available License v1.0. Website: www.qector.store.

QECTOR vs PyMatching

Circuit-level LER parity, latency gap disclosed

Real Stim circuit-level shots, rotated_memory_x, rounds = distance, p = 0.005, 40,000 shots per point through d=11. Same DEM decoded by both systems. QECTOR uses collapsed detector graph + weighted MWPM.

d	QECTOR-Blossom LER	PyMatching LER	QECTOR µs/shot	PyMatching µs/shot	Interpretation
3	0.0117	0.0117	0.7	0.5	LER parity
5	0.0089	0.0089	11.4	3.7	LER parity
7	0.0053	0.0053	60.4	13.3	LER parity
9	0.0029	0.0029	311.5	19.7	LER parity
11	0.0017	0.0017	633.9	46.0	LER parity

Honest latency note: PyMatching remains the latency leader for exact MWPM. QECTOR's value is reproducible workflow packaging, belief-matching accuracy mode, BP-OSD / LDPC support, and CPU/GPU batch workflows.

Extended Distances

LER parity extends to tested d=13 and d=15 workloads

The adaptive-k fix restored d=15 logical-error-rate parity. At larger distances QECTOR's matching weight path is best described as near-exact adaptive-k, not universally bit-exact MWPM; the tested logical outcomes remain at parity.

d	Shots	QECTOR LER	PyMatching LER	Errors Q/PM	QECTOR µs/shot
13	20,000	0.0010	0.0010	20 / 20	1437.3
15	20,000	0.0005	0.0005	10 / 10	3555.0

Belief-Matching

Lower observed LER on selected correlated workloads

Belief-matching runs BP on the hyperedge detector graph, preserving correlations that plain graphlike MWPM can discard, then reweights the matching problem. It is an accuracy mode and is slower than plain MWPM.

d	PyMatching LER	QECTOR belief LER	Reduction	Interpretation
3	0.0112	0.0109	2.7%	near parity at tiny distance
5	0.0084	0.0056	33.7%	headline advantage
7	0.0053	0.0039	25.7%	robust advantage

Multi-seed robustness

d=5, p=0.005, 8 independent seeds: belief-matching was better in 8/8 seeds with pooled LER reduction of 20.1%.

Seed × p grid

d=5, p=0.004/0.005/0.006, five seeds each: pooled reductions remain positive across the grid.

BP-OSD / LDPC

LDPC and qLDPC workflows beyond graphlike matching

Matching cannot decode codes whose mechanisms touch more than two detectors. QECTOR includes a BP-OSD workflow for arbitrary GF(2) check matrices and LDPC/qLDPC experiments.

Code	p	QECTOR BP-OSD LER	ldpc package LER	Result
[[72,12]] bivariate-bicycle	0.03	0.0370	0.0340	within ~10% of reference

GPU / Memory

CPU/GPU consistency and bounded memory profile

GPU correctness

CUDA and OpenCL batch decoders were checked against the CPU reference across d=3..13 and batches up to 65,536 on an NVIDIA GTX 1660 Ti. Outputs were bit-identical and syndrome-faithful in the tested configurations.

Memory and scaling

Python-side peak allocation remains approximately flat in the hot path across d=3..11, while native/Rust memory is measured separately through RSS profiling.

Reproduction

Recommended verification path

Build and test

git clone https://github.com/qectorlab/qector-decoder.git
cd qector-decoder
pip install maturin
maturin develop --release --no-default-features
pip install -e ".[dev]"
pytest

Benchmark evidence folders

Use the repository folders docs/, benchmark_results/, artifacts/, and scripts/ for methodology, results, hashes, and reproducibility drivers.

Limitations

Claims kept honest

Not a fastest-decoder claim.
PyMatching leads latency for exact MWPM on larger tested workloads.

Belief-matching is slower.
It lowers LER in selected correlated workloads but has per-shot rebuild cost.

Large-d MWPM wording matters.
Adaptive-k is near-exact at larger distances with LER parity in tested audits; do not market it as bit-exact everywhere.

Commercial use is not free.
QECTOR is source-available, not open-source. Commercial use requires a paid license.