HRR - Selective Logical-Sector Rescue for QLDPC Decoding
Kyuchul Lee, cording.ai
Abstract
HRR-HYBRID Under Stronger Validation: A Reproducible Correction and Negative Result for Selective Logical-Sector Rescue Abstract Version 1 reported that HRR-HYBRID improved several quantum low-density parity-check decoding baselines on asynthetic fixture. A subsequent audit identified design weaknesses: the check matrix was not a demonstratedquantum code family, errors were planted rather than sampled from the declared channel, train and test recordscould share exact error patterns, the random control had a smaller candidate budget, paired significance was notreported, and hybrid latency was not measured. We redesigned the experiment around three verifiedhypergraph-product CSS codes, separated fitting, validation, and testing by code size, used independent Bernoulli Xerrors, matched candidate budgets, and added paired statistical tests and end-to-end reference timing. On the held-out [[50,2,5]] code, HRR improved 7, tied 6, and worsened 7 of 20 decoder-by-error-rate cells.BeliefFind at physical error rate 0.10 showed an individual reduction in logical error rate from 0.280273 to 0.268066(48 fixes, 23 breaks; exact McNemar p=0.004065), but this did not survive a conservative correction across 20examined cells. Matched random candidates were better than HRR in two nearby BeliefFind cells. The sequentialreference hybrid was approximately 77-189 times slower than the corresponding baseline. These data do not supportthe v1 characterization of HRR-HYBRID as a generally effective, decoder-agnostic, deployment-oriented rescuelayer. They support continued study of a narrower, decoder-specific high-noise hypothesis. We publish the negativeresult, executable code, fixture, raw logs, dependency locks, and validation scripts as a correction to the publicrecord.