Most teams discover their inference latency in production. We write it into the design document before the first training run.

Budget first, model second

• P99 end-to-end: 40 ms, including network.
• Feature retrieval: 8 ms (streaming store, point-in-time correct).
• Model: 25 ms (quantised, batch-1 friendly).
• Overhead: 7 ms for tracing and SHAP capture.

Keeping the contract

A registry with signed artefacts, canary stages, and drift monitors that alert on the input distribution — not just the outcome. When a budget breaks in CI, the model does not ship.

The payoff is boring, and boring is the point: every decision reproducible, every percentile defensible.