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.
Status: pre-alpha. The plugin contract, the CLI, and the whole
translaterecipe exist and run end to end — Fortran in, verified NumPy and evidence manifests out, gated bit-exact against the compiled original. The other three recipes have their contracts and await their plugins (P3–P4) — seedocs/roadmap.md.
RecastEngine is not a translator. Translation is one recipe of four, and all four are the same five steps with different plugins in the slots:
discover -> analyze -> transform -> verify -> record
Unit Facts Candidate Verdict Evidence
| Recipe | What it does | Example product |
|---|---|---|
translate | Fortran → NumPy / Numba / CUDA | PyCAM5 |
refactor | carve a Python control plane into a Fortran monolith, numerics untouched | freeCAM |
port | retarget a kernel to an accelerator | JaxCAM6 |
audit | secret scan, SBOM+CVE+VEX, LLM source audit, sanitizer builds | CC-Test (cyber half, restricted access) |
uv sync --extra fortran --extra translate --extra verify --extra dev
uv run recast doctor # check the installation
uv run recast recipes # what workloads existThen run the whole translate recipe over the shipped example (needs a
gfortran on PATH — the reference really is compiled):
$ uv run recast run translate examples/toy_physics --config examples/toy_physics/recast.jsonfortran: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) recorded1 unit(s), 3 verdict(s), all passedOne command, eight stages: the module is translated, its dataflow
cross-checked against the source's, the untouched Fortran compiled as the
reference, every output compared bit for bit, and three evidence manifests
written under examples/toy_physics/.recast/evidence/ — because a candidate
without evidence is a draft, however good it looks. The generated Python
itself lands beside them under .recast/translate/.../candidate/.
recast plan dry-runs a recipe and reports what is missing before anything
costs compute — here, that the Numba backend is not migrated yet:
$ uv run recast plan translate --config '{"target": "numba"}' 1. [ok ] executor local 2. [ok ] frontend fortran 3. [MISS] transform translate.numba 4. [ok ] verifier static.rwset gate 5. [ok ] oracle f2py-golden 6. [ok ] verifier differential.bitexact gate 7. [ok ] verifier symbolic.notary optional 8. [ok ] store fs-evidenceThe core installs with zero dependencies and stays importable without a
compiler, a GPU, or a model provider; everything heavier is an extra
(fortran, translate, verify, numba, jax, agents).
To extend it, implement one of the ten interfaces in
src/recast/plugins/ and register an entry point —
see docs/writing-a-plugin.md.
docs/architecture.md | the spine, the ten interfaces, where the boundaries fall |
docs/writing-a-plugin.md | how to extend the engine |
docs/roadmap.md | phases P0–P6 |
examples/ | source trees the shipped recipes run over end to end |
conformance/ | what a plugin must satisfy, as a suite that runs |
| Bugs, features | issues in this repository |
| Vulnerabilities | SECURITY.md — private advisory, never a public issue |
| Collaboration, licensing | Yueqi Chen, University of Colorado Boulder — yueqi.chen@colorado.edu |
Contributing: CONTRIBUTING.md and AGENTS.md.
Improving the engine improves it for everyone using it — the plugin contract is
the same one every extension uses, so a plugin you write is not second-class.