Blog

What a research-to-production handover should contain

· 6 min read · 75bit

At some point a result stops being yours and becomes a system somebody else has to run. That moment is usually informal: a meeting, a repository link, a promise to answer questions. The people receiving it inherit everything the prototype assumed and nothing about why.

Production readiness can be written down

Google engineers turned this into a checklist. The ML Test Score is a rubric of 28 tests, drawn from running production machine learning systems for years, that scores how ready a system is to go live and how much technical debt it carries. You do not need to adopt it wholesale. It is useful as proof that readiness is a finite list of concrete checks, not a feeling.

28

concrete tests in Google's ML Test Score rubric for production readiness and technical debt reduction.

Breck et al., The ML Test Score.

A handover does not need 28 items. It needs a small set of things that are specific enough to act on.

The essentials

The receiving team needsIn practice
What it is forThe decision or product this supports, and the failure that would matter most
What it assumesInput ranges, data freshness, the conditions it was validated on, where it degrades
How to rebuild itOne command from a clean checkout, with versions pinned
How to know it is healthyThe few signals that matter, and the thresholds that should page someone
What to do when it is notA first-response runbook a person who did not build it can follow at 2 a.m.
Who to askNamed owners for the method, the data and the infrastructure, and how long support lasts

Test the handover

The only reliable check is to have the receiving team do something with it before the original team steps away: rebuild it, change a parameter, respond to a simulated incident. Every place they get stuck is a line missing from the document.

  1. Receiving team rebuilds the artefact from a clean checkout, without help.
  2. They change one parameter and explain what should happen before running it.
  3. They respond to a simulated failure using only the runbook.
  4. They list every question they had to ask. That list is the next version of the document.

Handovers fail quietly. The system keeps running until the first change nobody understands. A small amount of structure up front is much cheaper than rediscovering the assumptions later.

Sources