Scientific software redevelopment must preserve established numerical behavior. Genesis’s MESA evidence record, dated September 7, 2026, describes rewriting 13 Fortran modules as Rust crates, using the existing implementation as a reference.
The record reports 139,414 lines of Fortran → 89,946 lines of Rust, 1,052 tests passing, and US$10.64 in model-token charges. The repository documents golden tests against original Fortran outputs and reports 18 ignored tests alongside the passing tests.
The Evidence page gives 1.55–6.87× median speedups. These are reported benchmark figures, not a universal acceleration of MESA. Separately, the repository README’s August 5 benchmark table reports 1.25× for its end-to-end one-zone burn proxy. Different snapshots and workloads should remain distinct.
The port covers the selected numerical and physics modules, not the complete MESA system; stellar and binary evolution engines are outside the documented scope. The case shows how an existing codebase can guide redevelopment through behavior checks. Results are project-reported, without independent reproduction, and the token cost excludes hardware.