Team discussing ownership during an operations meeting

8 January 2026

Handing models to on-call without drama

Research teams often finish a model while platform engineers inherit undefined alerts. The first production week then becomes a negotiation about who wakes up for a drift page.

We ask for a one-page ownership matrix covering training jobs, serving endpoints, and monitoring. Each row needs a primary and a backup, plus a rollback path that does not require the original author.

That document is unglamorous. It is also the difference between a controlled pilot and a feature that gets switched off the first time metrics wobble.

Back to field notes