Lesson 046: Evaluation habits that compound
Focus
Bias toward observable metrics, not model marketing. Token Evaluation habits that compound:46 keeps neighbouring lessons differentiable.
Key ideas
- Thread: Evaluation habits that compound · drill v6 · spin
56007. - Habit: pair every model utterance with a trace_id you could paste into Grafana.
- Guardrail: write one RACI bullet referencing this lesson tomorrow.
Deep dive notebook
Synthetic drill artefacts
Red-team tableau
1. Actor profile **exec_assistant**
2. Injection meme `OVERRIDE++669`
3. Detector gate `moderation-tier-0` + human pager `oncall-ai-3`
4. Telemetry fields `trace_id,user_bucket,redaction_notes`
Practice
Practice Simulate degraded retrieval once; screenshot graceful degradation copy. — 46 Bump literals mindset by 32.