Google-validated quantum error-correction decoder backfires at target error rates

A preprint from Leiden University and Riverlane researchers shows the correlated-matching decoder used by Google and Riverlane degrades at physical error rates near 10⁻⁵, failing 4.4 times more often than plain MWPM. The error-suppression exponent drops from 1.98 to 1.28, making a distance-3 code behave like distance-2. The authors propose a single-parameter fix with no added latency.

• At 10⁻⁵ error rate the decoder fails 4.4x more often than uncorrelated MWPM
• Error-suppression exponent falls from 1.98 to 1.28
• Code distances 3 to 9 tested under the SI1000 noise model
• Failure reproduced in two independent production decoder implementations