Week 11 · Field delivery

Is the new model better?

0:00 / 3:00

15 seconds

Vendor said "better."

You flipped 100% traffic. Faithfulness dropped 9 points.

Model AControlModel BCandidateScoreWins
Lesson notesIs the new model better?: the short version+

Week 11 · Observability & break-glass · about 3 minutes

What you will be able to do

Run Helix's old and new models side by side and count wins before switching traffic.

The idea

Shadow or A/B the candidate on the golden set (and a live sample). Count wins on faithfulness, latency, cost.

Switch only when the numbers clear the bar.

wins = sum(score(b, q) > score(a, q) for q in suite)
assert wins / len(suite) >= 0.55

Marketing slides are not an A/B test.

Key takeaways

  • Side-by-side on a suite.
  • Track quality, latency, $.
  • Ramp on evidence.

Part of the Forward Deployed Engineer roadmap. New to the role? Start with what a Forward Deployed Engineer does.