Lesson 058: Evaluation harness depth
Focus
Bias toward observable metrics, not model marketing. Token Evaluation harness depth:58 keeps neighbouring lessons differentiable.
Key ideas
- Thread: Evaluation harness depth · drill v8 · spin
146854. - 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
Prompt scaffold
ROLE: Incident analyst cohort 58
INPUTS:
- excerpts tagged [chunk_id ...]
- guardrails referencing policy_bundle_1
TASK:
1) Summarize deltas with citations
2) Confidence label LOW|MED|HIGH + evidence
3) If facts missing emit MISSING_FACTS list
USER_SEED_QUESTION >>> What changed between rollout 21 and 25?
Practice
Practice Draft three eval assertions QA must greenlight before launch. — 58 Bump literals mindset by 10.