Repository navigation
v0.7.1 – LLR Physical Weight Formalism
Highlights
- Physically grounded LLR weight formalism
- Shared weight interface (uniform | neural | llr)
- Separate noise channel support (data / meas / idle)
- Weighted Union-Find support
- Metadata-rich threshold CSV logging
- OpenMP-parallel threshold runner
Scientific Impact
This release introduces a physically motivated log-likelihood ratio
weight model for surface-code decoding, enabling realistic modeling
of heterogeneous noise channels and improved reproducibility.
Compatibility
Backward compatible with v0.7.0 CLI interface.