Install

Python 3.10 or later. The package depends on numpy and pandas only. Pin the version in anything that has to be reproducible:

python -m pip install pynrpf==0.4.0

Databricks

Pin the version. Two routes give the same files, built from the same tagged commit by the release workflow:

1. PyPI, when the workspace can reach it: %pip install pynrpf==0.4.0 in a notebook, or pynrpf==0.4.0 in the cluster or job library list.

2. A wheel in a Unity Catalog Volume, when it cannot: download pynrpf-0.4.0-py3-none-any.whl and SHA256SUMS from the GitHub Release, check the sum, copy the wheel to a Volume and install it by path:

   %pip install /Volumes/<catalog>/<schema>/<volume>/pynrpf-0.4.0-py3-none-any.whl
   

The runtime needs Python 3.10 or later. pyspark is already on the cluster, so the Spark adapter needs nothing beyond the package. examples/databricks_smoke.py is a notebook that installs the pin and runs the method on synthetic data, in pandas and through Spark; run it once on a new cluster before any real readings.

Show details

From a clone of the repository

python -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install -e .

Extras

extra	adds	when you need it
fit	scikit-learn	fitting both calibration numbers; the intercept-only refit needs nothing extra
spark	pyspark	the Spark adapter, outside Databricks
paper	scikit-learn, xgboost, holidays, pyyaml, pyarrow, matplotlib, jupyterlab, tabulate	reproducing the journal paper
dev	pytest, ruff, build, mkdocs	working on the repository

python -m pip install -e .[dev,paper]

Reproducibility

The method has no randomness: the same readings, the same version and the same settings give the same tables. Pin the version, and keep the provenance record with every output: result.provenance in Python, provenance.json from the command line, and pynrpf.provenance(c, calibration, phi) for a Spark job. The record holds the package version, the two calibration numbers and their note, the evidence floor and the control.