Repository navigation
First release. An offline benchmark for routing, cited draft replies and abstention on a small fictional service-enquiry task: a zero-dependency Python package and CLI that scores prediction files you supply over a frozen 48-case synthetic set. It makes no model or network calls.
Assets: the wheel and source distribution built from this commit, with their SHA-256 sums in SHA256SUMS.txt. The same two files are published on PyPI as enquiry-bench.
Install: pip install enquiry-bench, or pip install "git+https://github.com/failstocode66/enquiry-bench@v0.1.0", then enquiry-bench demo --output demo-out. The demo figures (reference 48/48, always-abstain 12/48) are designed fixture results, not model performance.
Verified for this release: the 36-test suite passes on Windows with CPython 3.12.12, 3.13.12 and 3.14.5; the wheel's 10 runtime files are byte-identical to the wheel validated on 2026-11-03; the exact wheel installs offline into a fresh environment and reproduces the demo. See CHANGELOG.md for the history, including the earlier link-fixture refusal and what it does and does not show. Other platforms are unverified.