Heavy-hex-native bivariate bicycle codes from evolutionary and LLM-guided search: code, data, and verification artifacts

Complete replication package for the manuscript Heavy-hex-native bivariate bicycle codes from evolutionary and LLM-guided search by Hossain, Maji, Burnwal, and Taxak (2026). W. Hossain, S. Maji, and D. K. Burnwal contributed equally (co-first authors); S. Taxak is the corresponding author. Includes: the full search-and-evaluation pipeline (Python) with BP+OSD coset-decoding distance estimator, Smith-Normal-Form torus-equivalence and polynomial transport, MIP exact-distance solver (PuLP + CBC), heavy-hex locality metric, Stim-based circuit-level memory experiments; random, evolutionary, and LLM-guided search implementations with matched random controls; raw data (60,000-candidate random search CSV, evolutionary CSV, LLM round 1/2/3 proposal CSVs, circuit-level threshold Monte Carlo outputs, MIP exact-distance JSON certificates, deep BP+OSD verification logs for each top candidate); explicit polynomial pairs for the seven discovered heavy-hex-compliant BB codes; figure-generation scripts; LaTeX sources plus compiled PDFs; step-by-step reproducibility README with RNG seeds, wall-clock figures, software versions. All results reproducible on a single 12-core workstation in ~43 hours of wall-clock time.