Metadata-Version: 2.4
Name: exactref
Version: 0.0.1
Summary: Pure-Python reference implementations for checking numerical results (name reservation; in development)
License-Expression: MIT
Keywords: reference implementation,numerical accuracy,floating point,correct rounding,bfloat16,float8,oracle,testing
Classifier: Development Status :: 1 - Planning
Classifier: Intended Audience :: Developers
Classifier: Intended Audience :: Science/Research
Classifier: Operating System :: OS Independent
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3 :: Only
Classifier: Topic :: Scientific/Engineering :: Mathematics
Classifier: Topic :: Software Development :: Testing
Requires-Python: >=3.10
Description-Content-Type: text/markdown
License-File: LICENSE
Dynamic: license-file

# exactref

exactref is an early-stage project to write pure-Python reference implementations of mathematical operations: trustworthy answers that other numerical code can be checked against.

> **This release only reserves the package name.** It contains no functionality yet, so please don't depend on it.

## What we're aiming for

- **Portable:** runs on Python alone, with no NumPy, PyTorch, GPU, or native numerical libraries required.
- **Honest about accuracy:** each routine states the guarantee it actually provides (exact, correctly rounded, a certified bound, or an approximation) instead of a blanket "high precision" label.
- **Faithful to the data:** values are held in their real binary formats, including bfloat16 and float8, so rounding, signed zeros, infinities and NaNs are handled explicitly.

Correctness comes first and speed second. Coverage will start small and grow over time.

## License

MIT
