Benchmarks and methodology
Compare recorded sampling and decoding results, then inspect their conditions and reproduction steps.
Decoder results you can inspect
Compare logical failures and decoding time for the recorded surface-code and BB cases. Each result includes its conditions, source data, and reproduction instructions.
How to read the evidence
Smoke runs check wiring; checked full artifacts support only their named cases and recorded environments. A passing consistency check is not publication readiness or a general performance claim. Feature support is described separately in the support contract.
Circuit sampling
Setup time, sampling time and output costs on a named surface-code fixture.
Decoder comparisons
Logical-error-rate and time plots for the recorded surface and BB cases.
Detection and DEM sampling
Separate workloads and metrics; avoid comparing throughput with elapsed time.
Design a comparable measurement
Record the source revision, tool and dependency versions, CPU/OS, circuit and input hashes, noise channels and probabilities, distance, rounds, seed, shots, decoder configuration and stopping rule. Compare identical inputs and separate setup/compile, sampling, decoding, transport and output costs. Report failures and timeouts alongside successful shots; do not silently drop them from the denominator.
Logical failure rate is failures divided by completed scored shots. Include statistical intervals; zero failures is a finite-sample upper-bound result, not proof of zero risk. Timing needs warmups, repetitions and a stated aggregation; a single machine's median cannot establish a universal speedup. Full evidence below includes manifests, raw data and reproduction commands.
Historical measurements and regression checks
These records support investigation and compatibility checks. They are not the first-run tutorial or a current universal performance ranking.
Loading historical evidence.
Local workflow checks
These local-only outputs test setup, readiness and workflow behavior. Follow the linked repository instructions to reproduce them.
Loading local workflow evidence.
88.02% smaller than raw b8
Pinned surface-code case: d=11, 100 rounds, 1,024 shots.
Surface code d=11, rounds=100
1,024 shots| Format | Bytes | Share of raw |
|---|---|---|
| Raw b8 | 1,552,384 | 100.00% |
| Direct Zstandard | 325,565 | 20.97% |
| RSMP v1 | 186,028 | 11.98% |
Source: RSMP evidence bundle. Keep archive encoding, circuit and sample size fixed when reproducing; compression is separate from simulation and decoding throughput.