The modernization engine of SciRecast.

LLM agents do the labor of modernizing legacy scientific software; RecastEngine is the part that is reusable across every such effort, and the part that refuses to let anything through that has not been checked against the original.

Python 3.11 or newer. Activate the environment once and every command after that is a bare recast.

uv venv --python 3.11                              # fetches a 3.11 if you have none
source .venv/bin/activate                          # Windows: .venv\Scripts\activate
uv pip install -e ".[fortran,translate,verify]"

Without uv: python3 -m venv .venv and pip install -e with the same extras. The core itself has zero dependencies and stays importable without a compiler, a GPU, or a model provider — fortran, translate, verify, numba, jax, agents and all are the extras.

$ recast doctor
recast 0.0.1.dev0  python 3.11.16
23 plugin(s) registered across 10 kinds

The shipped example, all the way through. Needs a gfortran on PATH — the reference really is compiled:

$ recast run translate examples/toy_physics --config examples/toy_physics/recast.json
fortran:toy_physics
  [ok ] frontend   fortran
  [ok ] transform  translate.numpy
  [ok ] verifier   static.rwset                sampled: 4 blocks match
  [ok ] oracle     f2py-golden                 f2py:toy_physics:3f3e0f78...
  [ok ] verifier   differential.bitexact       bit_exact: 85 points across 2 subprogram(s), all bit-exact
  [ok ] verifier   symbolic.notary             symbolic: no rewrites to notarize; the translation is print-order faithful
  [ok ] store      fs-evidence                 3 verdict(s) recorded

1 unit(s), 3 verdict(s), all passed

One command, eight stages: the module translated, its dataflow cross-checked against the source's, the untouched Fortran compiled as the reference, every output compared bit for bit. Nothing is written into the source tree — it all lands in output/, under the project's own name:

toy_physics is the project's name; fortran_toy_physics is the one unit it declares — the fortran:toy_physics printed at the top of the run, with the colon made path-safe. A tree that declares several units gets one such directory each.

For the same thing on code this project did not write, corpus/ pins twelve open-source Fortran libraries and runs the engine over them with no domain extension installed. docs/corpus-numfor-example.md walks one of those units through the same eight stages, and says how far its passing run reaches.

The four recipes, the config keys, the flags on run and what its exit codes mean: docs/cli.md.

Apache License 2.0 — see LICENSE and NOTICE.