Skip to content

CONTRIBUTING.md

Latest commit

 

History

History
50 lines (41 loc) · 2.11 KB

CONTRIBUTING.md

File metadata and controls

50 lines (41 loc) · 2.11 KB
Raw

Contributing to skaters

Thanks for considering a contribution. The package has strict invariants; this page tells you what they are so your change passes review on the first try.

Setup and tests

git clone https://github.com/microprediction/skaters
cd skaters
pip install -e ".[dev]"
pytest

The suite must pass in full. It includes three kinds of tests beyond the unit tests, and all three gate every release:

  1. JS parity (tests/test_js_parity.py): the JavaScript twin in docs/js/skaters/ must agree with Python to 1e-6 across ~100,000 probe values. If you change numerical behavior in one language, change the other, then regenerate the vectors: PYTHONPATH=src python parity/gen_vectors.py && node parity/check.mjs.
  2. Adversarial gates (tests/test_tails_robustness.py, parity/adversarial.mjs): the deployed default must survive constant series, lattices, monster spikes, scale collapse, and vol whiplash, and recover afterwards. If you find a new failure mode in the wild, add the stream to these files as part of the fix.
  3. The laplace contract (tests/test_laplace_contract.py): the signature, state surface, determinism, and checkpoint-restore equivalence of laplace are frozen. People deploy this function. Changing its contract requires editing that test deliberately, plus a changelog entry and a version bump, in the same pull request.

Ground rules

  • Pure stdlib in src/skaters/; dependencies belong in benchmarks only.
  • Skater state is pure data: picklable, resumable, no closures.
  • Strictly causal everywhere: score before update; nothing may read the future or whole-series statistics.
  • New estimators forget: prefer EWMA-style memory to cumulative counts.
  • Benchmark claims come with committed study files. If you assert a number, commit the harness and results that produce it (see benchmarks/anomaly/RESULTS.md for the format).

Reporting issues

Open a GitHub issue with a minimal reproducing stream where possible. Calibration complaints are especially welcome: a PIT histogram from a quiet stretch of your data says more than an adjective.