Versioned datasets

Three decoders (pymatching MWPM, beliefmatching, tesseract) on byte-identical Stim detector samples per configuration, rotated surface code, circuit-level depolarizing noise, 5000 shared shots, seed 7. Measured configurations: d=3 at p=0.008 and p=0.02, and d=5 at p=0.008. d=5 at p=0.02 and all d=7 cells are ABSENT, not zero: the tesseract timing child dies there (suspected OOM in the most-likely-error search; ketqat-benchmarks#8) and an absent measurement is recorded as absence. Accuracy cells are measured counts on the shared sample; latency percentiles (500 shots x 3 randomized repeats), batch throughput and peak RSS are measured per decoder in an isolated child process on the machine named in each row and do not transfer to other machines. Pauli/Stim regime only.

3 versions

Per-run rows for the threshold sweep behind the rotated-surface-code family page: distances 3, 5, 7 and 9 at six physical error rates bracketing threshold (0.002-0.02), three seeds per configuration, shots tiered 20k/100k/200k by failure rarity. Each row is copied verbatim from one hash-verified registry run named in the lineage. Seeds are pooled downstream by summing failures and shots; this table keeps them separate so the pooling is checkable.

1 version

Per-configuration logical error rates for the pymatching MWPM baseline on the rotated surface code under circuit-level depolarizing noise: distances 3, 5 and 7 at eight physical error rates each, 10,000 shots per configuration. Each row is derived from exactly one hash-verified registry run named in the lineage. Rows with zero observed logical failures report a rate of 0 with a null standard error; the source runs carry the Wilson upper bound.

1 version