Quantify the robustness of optimized decoder priors under hardware drift

Determine which features of decoder-prior optimization are responsible for the apparent stability of the IBM LEP-optimized decoder and Google's reinforcement-learning-optimized prior under hardware variation, and jointly characterize decoder-prior optimization, temporal robustness, and likelihood-aware confidence under drift.

Background

The paper compares calibration-derived and hardware-optimized decoder priors on IBM and Google quantum-memory data. It reports suggestive evidence that priors optimized directly for logical performance are more robust to hardware variation, while also emphasizing that the observed stability was not quantified and that its causal features were not identified.

The unresolved problem is therefore to establish whether decoder-prior optimization systematically improves robustness to temporal and distributional noise drift, identify which aspects of the optimization produce that robustness, and study its interaction with likelihood-aware postselection confidence.