Production is the only benchmark that pays: monitor, drift, feedback
Input and output drift, per-group outcomes, heartbeats and the retrain trigger
The question: Your system passed its evals on launch day. How will you know, in June, that it is still doing what it did in March?
Models are the only asset that gets worse while sitting still: the world drifts away from what they learned. The eval harness proves a system at a moment; production monitoring proves it over time — and it is where the silent failures live.
What the lesson covers
Drift has two faces. **Input drift**: the cases arriving in production stop looking like the cases the system was built and evaluated on — new products, new phrasing, a new customer segment, a renamed field. **Output drift**: the system's answers change in distribution — more refusals, longer replies, a class that is suddenly never predicted. Either can happen with no error raised, and both are invisible unless something is watching.
The monitoring stack is unglamorous and decisive. A **sample** of production cases graded on a cadence with the Lesson 39 rubric — the eval harness pointed at live traffic. **Distribution checks** on inputs and outputs against the launch baseline. **Per-group outcomes** on the groups you named in the fairness memo (Lesson 17), because fairness decays too. A **zero-runs heartbeat** and an error path to a human (Lesson 22). And an owner who is paged — the three-in-the-morning question from Lesson 7.
**Feedback loops** are the second half. Overrides, edits, escalations and complaints are the richest signal a production system produces; logging them is cheap, reading them is where improvement comes from, and adding them to the eval set (Lesson 39) is how the harness learns what the world taught you. A system with no feedback path is a system whose failures are discovered by customers.
Define the **retrain or rollback trigger** in advance, as a number: sampled quality below X, a drift statistic beyond Y, a per-group gap above Z — and the action each triggers. A trigger defined after the fact is a negotiation; one defined before is a policy. This is NIST AI RMF's Measure and Manage functions made concrete for one workflow.
The economics (Lesson 31) are in this lesson too: monitoring is the maintenance line nobody prices, and the cheapest time to fund it is before launch, when it is a design choice rather than an incident response.
Key points
- Drift has two faces — input and output — and both arrive silently.
- Six standing checks: graded sample, distribution, per-group outcomes, heartbeat, feedback path, a paged owner.
- Overrides and complaints are the richest signal; log them and add them to the eval set.
- Define the retrain/rollback trigger as a number before launch; afterwards it is a negotiation.
- Monitoring is the maintenance line nobody prices — fund it as design, not as incident response.
Framework — Sample → Distribution → Per-group → Heartbeat → Feedback → Trigger
Six standing checks on a live system, each with an owner, a cadence and a number that triggers an action.
The lab
Design and, where possible, instrument production monitoring for one live or planned AI workflow.
Open this lesson, its lab and its quiz
Sources and further reading
- AI Risk Management Framework — Measure and Manage — NIST
- Hidden Technical Debt in Machine Learning Systems — Sculley et al., Google
- Lesson 22 — the heartbeat and the silent failure — Applied AI Academy