Tutorial 3: Wire CI/CD quality gates

Tutorial 3 — Wire Stratix evaluations into GitHub Actions to gate AI changes.

Time: ~30 minutes Level: Intermediate You'll build: A GitHub Actions workflow that runs a Stratix evaluation on every PR and blocks merge on regression.

What you'll learn

• How to call Stratix from a CI runner

• How to compare a run against a baseline

• How to fail a build on regression

Prerequisites

• Completed Tutorials 1 and 2

• A GitHub repo with an AI feature you can iterate on

• A Stratix evaluation space you want to gate against

• A GitHub Actions runner

Step 1: Get an API key

In Stratix Premium → Settings → API keys → New key. Scope it to evaluation read/write. Copy the key.

In GitHub: Settings → Secrets and variables → Actions → New repository secret. Name: LAYERLENS_STRATIX_API_KEY.

Step 2: Create the CI script

scripts/ci-eval.py:

import os
import sys
from layerlens import Stratix

client = Stratix(api_key=os.environ["LAYERLENS_STRATIX_API_KEY"])

# Re-run the gating space
space_id = "your-space-id"
run = client.spaces.run(space_id=space_id)
result = client.spaces.wait_for_completion(run.id, timeout=600)

# Fetch baseline (most recent main-branch run)
baseline = client.spaces.latest_run(space_id=space_id, branch="main")

# Compare
delta = result.score - baseline.score
tolerance = 0.01
print(f"Score: {result.score}")
print(f"Baseline: {baseline.score}")
print(f"Delta: {delta:.4f}")

if delta < -tolerance:
 print("REGRESSION beyond tolerance — failing the build.")
 sys.exit(1)
else:
 print("OK")

Step 3: Add the workflow

.github/workflows/eval.yml:

name: AI Eval Gate

on:
 pull_request:
 paths:
 - 'src/prompts/**'
 - 'src/agents/**'

jobs:
 eval:
 runs-on: ubuntu-latest
 steps:
 - uses: actions/checkout@v4
 - uses: actions/setup-python@v5
 with:
 python-version: '3.11'
 - name: Install
 run: |
 pip install layerlens --extra-index-url https://sdk.layerlens.ai/package
 - name: Run eval
 run: python scripts/ci-eval.py
 env:
 LAYERLENS_STRATIX_API_KEY: ${{ secrets.LAYERLENS_STRATIX_API_KEY }}

Step 4: Open a test PR

Make a small change to a prompt under src/prompts/. Push. Watch the GitHub Action run. Verify the eval kicks off.

Step 5: Tune the tolerance

Start strict (0.01). Loosen as you learn the natural variance of the evaluation. If you get false-positives on noise, you have two choices:

• Increase tolerance

• Make the eval less noisy (more rows, or a determined seed)

What's next

• Tutorial 4: Score live traces

• Use case: AI quality gates in CI/CD

• Workflow: Govern

• Integrations: GitHub Actions