Mistral describes a 40,000-line Fortran 77 to C++ migration built around a numerical parity harness

Mistral's post describes a migration of 40,000 lines of Fortran 77 to C++ at a European energy operator. The system is described as a physics-intensive reservoir simulator with no test suite and no centralized documentation — the kind of codebase where, as the post puts it, knowledge embedded in the code becomes hard to recover once the original authors leave.

The write-up treats translation as the easy part. "Translating syntax from one language to another is a largely solved task," the post states, and locates the difficulty in the surrounding system. Three elements are named as the work: building a parity harness for numerical verification, documenting the codebase with AI agents, and using structured workflows with human oversight. The lessons the authors draw are prioritizing numerical agreement, organizing documentation before migration, and balancing agent autonomy against human review.

For this desk this is an implementation note, not a tool announcement. Nothing in the evidence identifies which models or agent tooling were involved, what the parity harness actually compared, how long the work took, what the operator paid, or whether the migrated code is in production. That gap matters: the reusable idea here is the ordering — document first, migrate second — and the insistence that a rewrite of numerical code be checked against the old system's outputs rather than against a reviewer's impression. Both are workflow decisions a reader can adopt without any new purchase.

Readers who maintain long-lived scientific or numerical code should treat the parity harness as the transferable part and wait for method detail before copying the rest. The precise unresolved question is what the harness compared, at what tolerance, and who adjudicated the disagreements it surfaced.