Per-shot decoder confidences for "Soft information from neural-network logits and correlation-aware matching gaps"

Signed confidence of three decoders on the same sampled syndromes of rotated-surface-code memory experiments under uniform circuit-level noise (stim), 10^8 shots per run, for 20 runs: (d, r) = (5, 5), (5, 7), (5, 9), (5, 11), (7, 7), (7, 9), (7, 11), (9, 9) at p = 0.005, and d = r = 5, 7, 9 at p = 0.003, 0.0035, 0.004, 0.0045. Decoders: uncorrelated minimum-weight perfect matching (complementary gap), correlation-aware matching (complementary gap on a shared reweighted graph), and the pretrained graph-neural-network decoder of Lange et al., Phys. Rev. Research 7, 023181 (2025) (logit). Each run has one float32 NumPy array per decoder (positive where the decoder is correct, magnitude = confidence in dB) and a JSON file with the configuration, software versions and consistency checks. The code that generated the data and rebuilds every figure, table and number of the paper from it is at https://github.com/DavidDentelski/GNN-confidence-score. See README.md for the format.