Qector Sinter

QECTOR Sinter

Source of authority: v1.0.0 reference manual, section 17.2.

The sinter surface

from qector_decoder_v3 import qector_sinter_decoders
decoders = qector_sinter_decoders()
# -> {
#      "qector_blossom": <sinter.Decoder>,
#      "qector_belief": <sinter.Decoder>,
#      "qector_unionfind": <sinter.Decoder>,
#      "qector_bposd": <sinter.Decoder>,
#      "qector_unionfind_unweighted": <sinter.Decoder>,
#    }

These are sinter.Decoder entry points, so logical-error rates can be compared head-to-head in the community-standard harness with no QECTOR-specific tooling.

How to use them

import sinter
from qector_decoder_v3 import qector_sinter_decoders

decoders = qector_sinter_decoders()
# Run sinter with these decoders the same way you would any
# other sinter.Decoder:
results = sinter.collect(
    num_workers=4,
    tasks=[...],
    decoders=[decoders["qector_blossom"], decoders["qector_unionfind"]],
    max_shots=10_000,
)

The output is the standard sinter Stat collection, which you score with sinter.fit_binomial or with the QECTOR qector-research.wilson_ci for a 95% Wilson interval.

What each entry point does

• qector_blossom -> exact Blossom (MWPM), weighted by log((1-p)/p).
• qector_belief -> belief-matching: reweight the matching graph with BP posteriors per shot.
• qector_unionfind -> Union-Find with weighted growth.
• qector_unionfind_unweighted -> Union-Find unweighted; useful for code_capacity baselines.
• qector_bposd -> BP-OSD; for non-graphlike inputs.

How to probe

qector-research.sinter_decoder_list returns:

• sinter_exposed (bool)
• sinter_decoders (list of {"name", "type"})
• a note if qector_sinter_decoders is absent from the installed wheel.

The shipped qector_decoder_v3 1.0.0 wheel exposes the function; older or source builds may not. Always probe on the target device before publishing a sinter-based result.

Reference manual quote

qector_sinter_decoders() exposes qector_blossom, qector_belief, qector_unionfind, qector_bposd, and qector_unionfind_unweighted as sinter.Decoder entry points, so logical-error rates can be compared head-to-head in the community-standard harness with no QECTOR-specific tooling.

Common pitfalls

• Comparing sinter results across noise models -> refused by the competitive harness (manual 15.3). Tag the run.
• Comparing sinter results without a Wilson interval -> manual 15.2 requires a 95% Wilson interval for any published LER.
• Asserting "QECTOR sinter is fastest" without an artifact -> per-machine, per-workload only (manual 22.5).
• Mixing qector_blossom weighted and unweighted without flagging the difference -> the weight is in the DEM, not in the sinter entry point; the entry point always reads weights from the DEM.
• Using qector_bposd on a graphlike code -> it works, but the matching decoders are faster on tested configurations.

How the bench server helps

• qector-research.sinter_decoder_list is the live probe.
• qector-research.wilson_ci and wilson_table are the math utilities for the LER output.
• qector-research.logical_coset_score scores a batch of (predicted, sampled) logical observables on the logical coset (Theorem 2).
• qector-research.artifact_metadata_check generates the chapter 22.3 metadata block for a sinter-based artifact.