Phase 7: Production Observability & Incident Response · 45 min · Langfuse · Arize Phoenix · Python

The Four Incident Classes

Classify before you act. A PII leak and a subtle quality regression are both 'incidents,' but they call for opposite first moves.

Hiring signal: Correctly triaging an incident into the right class and severity tier before acting — instead of treating every alert identically — is exactly the on-call judgment a production AI system needs and most postmortems reveal was missing.

What you will learn

Name and distinguish the four incident classes: hallucination, jailbreak, drift, PII leak
Classify an incident's signals into the correct class before deciding a response
Apply the severity rubric: S1 (safety/regulatory, contain first), S2 (quality regression, RCA first), S3 (edge case, defer)
Explain why containment and root-cause analysis are ordered differently depending on incident class

Introduction

The Four Incident Classes

An on-call engineer gets two alerts within an hour of each other. The first: a sampled trace shows the model disclosed a customer's full account number in response to an unrelated question. The second: the rolling-mean quality score for the billing-questions route has drifted down 4 points over the last 35 minutes, with no deploy in that window. Both are real incidents. Treating them the same way — investigate root cause first, then decide what to do — is the wrong move for the first one and roughly right for the second. The first is a live privacy exposure that needs to stop happening right now, before anyone fully understands why; the second is a real but non-urgent quality problem where understanding the cause is exactly the right first step, because there's no active harm accumulating minute by minute while you investigate.

Four classes, four different signatures

Hallucination: the model's factual grounding drops — it states something false or unsupported as if it were fact, the production version of the poisoning failure mode from c12-02-3, now showing up as a live quality problem instead of a context-engineering exercise. Jailbreak: instruction adherence or privacy protections drop because someone successfully got the model to bypass its guardrails — a security-relevant class, distinct from an honest mistake. Drift: the rolling-mean quality signal from the previous lesson drops with no corresponding code deploy — meaning something changed in the environment around the system, not something the team shipped: a retrieved data source went stale, an upstream API's behavior shifted, a model provider silently updated something. PII leak: a privacy or safety score hits a near-bottom reading — personal or sensitive information exposed that shouldn't have been.

Why the class matters more than "something's wrong"
Each class points at a different likely cause and a different appropriate first move. A hallucination often traces back to a context or retrieval problem. A jailbreak traces back to a guardrail gap someone actively exploited. Drift with no deploy points outside your own codebase entirely. A PII leak is, regardless of cause, an active harm that needs to stop immediately. Classifying first is what tells you which of these very different investigations to actually start.

Unlock the full lesson

You've read the first 2 sections. The rest of this lesson covers Severity: contain first, or understand first, Jailbreak's severity, specifically, Classification errors compound, Build It — plus a hands-on lab, quiz, and project artifact.

Create a free account to unlock Phase 0 and Phase 1 of every course — no credit card.

Key terms in this lesson

Hallucination