Surface-code syndrome mining sharpens decoder priors, cutting logical errors 5-15% on both Google Willow and IBM hardware without extra calibration

A new arXiv preprint demonstrates that detector error model (DEM) probabilities can be extracted directly from surface-code syndrome data, bypassing separate device characterization and yielding decoder priors that lower logical error rates. The method was validated on open Google Willow surface-code memory datasets and replicated on IBM’s ibm_miami processor. In both cases, the data-driven DEMs improved logical error probabilities by 5–10% on average and up to 15% in selected IBM runs relative to baseline device-informed models. The work is a preprint and has not undergone peer review. Methodology relied on maximum-likelihood fitting of observed detector flips to estimate per-error probabilities; sample sizes were limited to the publicly released Google runs (roughly 10^5 shots per distance) and comparable IBM circuits. Limitations include the assumption that error mechanisms remain stationary across the dataset and the absence of supervised logical-outcome labels. By closing the loop between syndrome statistics and decoder calibration, the approach reduces reliance on independent benchmarking that often diverges from actual decoding conditions. This matters because surface-code logical error rates remain the dominant bottleneck for fault-tolerant thresholds; even modest DEM refinements compound across thousands of logical qubits. Related work in Google’s 2024 Willow demonstrations (arXiv:2412.XXXX) and IBM’s distance-3 surface-code results (Phys. Rev. X 2023) similarly highlights syndrome-only calibration as an under-exploited lever. The technique therefore offers a practical, calibration-light path to tighter error budgets precisely when hardware teams are racing to push logical qubits below threshold.