Skip to main content
OpsRoadmaps

Checklists / model-production-readiness

Model Production Readiness Checklist

Before a model influences a real decision: reproducible training, consistent features, an evaluation gate, drift monitoring, and a rollback someone has actually performed.

9 steps · 8 required · last reviewed (next review 2026-12-20) · updated 2026-09-20

Related focus areas:MLOps

Sign in to run this checklist

Run this before a model serves production traffic, and again whenever the model's consequence increases. The distinguishing risk of a machine-learning system is that it degrades without anyone changing it, while every infrastructure dashboard stays green.

If an item cannot be checked, that is a decision being made implicitly. Write down which ones you are consciously accepting.

  1. The training run behind this model can be reproduced from recorded inputs

    required

    A result you cannot reproduce is a result you cannot defend, debug or improve on.

  2. Features are computed by the same logic in training and serving

    required

    Training/serving skew is the most expensive recurring defect in production machine learning, and it is silent.

  3. An evaluation gate compares the candidate with the incumbent and can block the release

    required

    Evaluation that reports into a dashboard nobody gates on is measurement without consequence.

  4. Input distribution and prediction monitoring exist with defined thresholds and an owner

    required

    Nothing in your infrastructure monitoring will notice that the model became wrong.

  5. Rollback to the previous model version has been performed, not merely designed

    required

    The first execution of a rollback should not happen while a bad model is serving users.

  6. The model registry records which version is in production and what it was approved for

    required

    "Which model is live?" needing a person to answer is how two teams ship conflicting changes.

  7. A documented threshold exists at which the model is withdrawn from service

    required

    Deciding the pull criteria while the model is misbehaving guarantees it is decided slowly.

  8. Someone outside the building team can explain what the model is used for

    required

    A model whose purpose lives only with its authors cannot be governed and will outlive their tenure.

  9. Retraining is on a decided cadence or trigger rather than on request

    optional

    Retraining when someone remembers means retraining after the complaint.

Run it against your real gaps

A checklist is most useful when you know which controls you are actually missing. A free OpsRoadmaps assessment scores your controls with explainable findings and builds a roadmap from the results.

Assess my stack

Checklists document self-reported practice. Items were last reviewed 2026-09-20 — adapt them to your own environment.