Key Takeaway
Model review meetings that follow a consistent agenda ensure no critical production readiness check is skipped, reducing the risk of model incidents after deployment.
When to Use This Template
Use this agenda as the gate review before any model goes to production or receives a significant update. The review is typically attended by the presenting ML team, the platform/infrastructure team, the product owner, and a senior engineering leader with approval authority. Budget 60-90 minutes depending on model complexity.
Meeting Flow
- 1
Model Overview (10 min)
Presenting team summarizes: what the model does, what business problem it solves, architecture and training approach, data sources used, and what has changed since last review (if this is an update). Reviewers should have received a written summary as a pre-read.
- 2
Evaluation Results (15 min)
Present benchmark performance against baseline model and/or human performance. Show slice-based analysis for key user segments. Present fairness metrics across protected attributes. Show regression test results confirming no degradation on existing capabilities.
- 3
Production Readiness (15 min)
Present infrastructure requirements (compute, memory, storage), latency benchmarks under realistic load, scalability testing results at projected traffic, cost projections for serving at scale, and any infrastructure changes needed for deployment.
- 4
Monitoring and Deployment Plan (10 min)
Present the monitoring plan: which metrics will be tracked, alert thresholds, drift detection approach, and on-call runbook. Present the deployment strategy: rollout plan (canary percentage, duration), rollback criteria (automatic and manual triggers), and user communication plan.
- 5
Risk Assessment and Decision (10 min)
Review identified risks and mitigation measures. Decide: approve for deployment, approve with conditions (specific requirements that must be met before deployment), or defer with requirements (changes needed before the next review). Document the decision with rationale.
Facilitation Guidance
The review chair should ensure that discussion focuses on production readiness, not research quality. A model can be scientifically interesting but not production-ready, and vice versa. Common reasons for deferral include: insufficient evaluation on underrepresented user segments, missing monitoring for a critical quality dimension, and latency that does not meet the SLA under peak load. Always document the specific conditions for re-review when deferring.
Create a model review checklist that the presenting team completes before the review meeting. This reduces meeting time spent on items that could be verified asynchronously and focuses the meeting on items that require discussion and judgment.
Version History
1.0.0 · 2026-03-01
- • Initial model review agenda template