Pathfinder: Direction-Aware Neural Decoding and Complementary-Failure Ensembles for Surface Codes

I present two open-source decoder systems for quantum error correction on rotated surface codes, both built on Pathfinder, a direction-specific 3D CNN trained with the Muon optimizer. System 1 (canonical Pathfinder + a custom Triton kernel): a single 500K-parameter checkpoint per code distance reaches 6.43 us/syndrome at d=7, batch=1024 on an NVIDIA H200, batched throughput below a 7-us-per-syndrome service-rate target at every operational noise rate, while batch-1 latency (215 us) remains far above streaming real-time requirements. On accuracy it is never statistically beaten by PyMatching across 24 (distance, noise) points under 3-parameter circuit-level noise (20 point-estimate wins, 12 with non-overlapping 95% Wilson CIs), using one fixed, validation-selected checkpoint per distance; under the 4-parameter noise model of Lange et al. it is statistically indistinguishable from PyMatching and trails Lange's GNN by ~14% relative. System 2 (Pathfinder-Triad): a 3-way majority vote of PFWL3S (a wider, 3-seed, distillation-trained Pathfinder variant), Lange's GNN, and PyMatching exploits the decoders' complementary failure modes to reach 2.384% logical error rate at d=7, p=0.007 (100K shots), below each component decoder, with non-overlapping CIs versus published Lange, and strict-CI wins at five (d,p) points across d=7 and d=9. In the strongest controlled comparison (designated post hoc as primary), standalone PFWL3S beats an equal-size three-seed fine-tuned-Lange ensemble at d=7, p=0.010 under the paired McNemar test, surviving Holm and whole-paper Bonferroni correction. All accuracy results are simulation benchmarks on memory experiments; on real IBM Heron r2 data, decoded offline (outside the control loop), PFWL3S shows no statistically resolved difference from PyMatching (a null, not parity). Six negative results are documented. Code, checkpoints, and raw data are released. This record archives v2 of the preprint (v1: 2026-07-15). v2 addresses external engineering review: the headline latency re-measured on real-pipeline input shapes ([B,1,8,8,8]; 6.12 -> 6.43 us/syn, the sub-7-us service-rate claim survives with reduced margin), IBM results reframed as offline decoding of recorded measurement data, decoder-ensemble prior art added (Sheth et al. 2020; Shutty et al. 2026), and the paper restructured as a short main text plus supplementary material with a section crosswalk. Thanks to Joel Pendleton for the latency-shape catch. Code, trained checkpoints, and raw evaluation data: https://github.com/bledden/pathfinder. Feedback welcome; revisions will be published as new versions of this record.