Getting started¶
Install¶
pip install driftwatch-lifetime # or: uv tool install driftwatch-lifetime
driftwatch --version
Python 3.12 is required. The distribution name is driftwatch-lifetime and the import name is driftwatch. The unrelated PyPI package driftwatch (a general machine-learning drift library) installs the same import name, so do not install both in one environment (ADR-0003).
Check that it works (no data needed)¶
driftwatch smoke
smoke regenerates a small synthetic fixture (random values with the real file layout), runs the full pipeline on it (cache build, time slicer with the C1 to C3 checks, two detectors, A-AUT) and compares the table with a golden copy shipped in the package. It ends with a line starting MATCH and exit code 0. Fixture numbers test the code path only; they are never evidence.
Reproduce Table 1 (needs a clone and about 265 MB of data)¶
git clone https://github.com/Yathindra25KB/driftwatch-lifetime
cd driftwatch-lifetime
uv sync --frozen
uv run driftwatch reproduce table1
This fetches BODMAS, runs the E0 baseline and compares with the stored Table 1. On a clean clone the time-aware columns matched to four decimals and the random-split comparator columns within 0.006 (see the claims ledger, row E-24). Exit code 3 means the dataset host refused the download; the data card documents the manual path.
Rebuild the tables of E2 to E4¶
The tables and figures are rebuilt from stored per-slot predictions with no model fits (scripts/make_tables.py, make_e3.py, make_e4.py). The predictions are not committed; to rerun an experiment, build the data cache and follow the grid guide.
Where to go next¶
What was measured, with every number traceable: claims ledger, report chapters 7 and 8.
How the code enforces the protocol: architecture.
Using the code in your own study: the API reference.