CI/CD integration

A reference artifact lives in your repository (via git-LFS or a cloud bucket). Every PR re-records the same experiment and asserts a fingerprint match. When the two disagree, the job fails with a structured diff — never a mystery.

Goal

Fail the PR when a change to the model, the training script, or the pinned dependencies would silently alter recorded activations.

The workflow

name: neuronscope-gate
on: [pull_request]
jobs:
  gate:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with:
          lfs: true
      - uses: actions/setup-python@v5
        with:
          python-version: '3.12'
      - run: pip install "neuronscope[torch]==0.9.2" -r requirements.txt
      - run: python scripts/record_reference.py --out current.nsz
      - run: ns compare tests/fixtures/reference.nsz current.nsz --strict

Assertions

1. ns compare --strict exits non-zero on any diff above the metric's noise floor.
2. For per-layer thresholds, use ns compare --report compare.md and post the file as a PR comment.
3. To assert exactly-equal artifacts (bit-identical bytes on the same CPU/GPU), pass --bit-exact.

Operational notes