numguard: numerical edge-case testing with an optional LLM repair agent

Research applicability

Helps researchers inspect numerical correctness of scientific or ML array code, including AI-generated implementations.

Getting started

Editorial, untested guidance: read the README, install numguard-ci in an isolated environment, and test one non-sensitive numerical function with a trusted reference before enabling the optional agent.

Maintenance

Source availability checked on access date; future maintenance not established.

License

Licensing and redistribution terms must be checked before reuse; no independent licence audit.

Workflow & human oversight

Resource-level workflow is unspecified. Researchers set objectives, validate assumptions and outputs, and supervise any adoption.

Workflow level is not an evidence score; higher is not better evidence.

Approach

Fixed checks assess outputs against references and dtype limits; the LLM proposes a spec or rewrite rather than deciding correctness. Documentation describes CLI, pytest and CI integration.

Limitations

Partial repository documentation review only; no installation, security audit or outcome reproduction. The experimental agent success rate is unmeasured. Finite cases do not prove correctness; permissive specifications can hide bugs, and real optional backends are not exercised in the stated CI tests.

Reproducibility

Source review only; no code, hardware or claimed outcome reproduced.

Adoption requirements

Python 3.10+, NumPy and click; suitable functions, reference implementations and researcher-reviewed specifications. Optional model access entails API costs and sending code to a provider: review confidentiality and institutional policy first.