Lesson 159: Benchmarks read with scepticism
Focus
Assume an auditor reruns everything you claim; narrate checkpoints aloud. Token Benchmarks read with scepticism:159 keeps neighbouring lessons differentiable.
Key ideas
- Thread: Benchmarks read with scepticism · drill v9 · spin
196286. - 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
Retrieval partitioning plan
| Slice | Tokens | Retrieval mode | Notes |
|-------|--------|----------------|-------|
| FAQs | 502 | hybrid@0.56 | keep tables contiguous |
| Policies | 585 | dense@0.74 | include footnotes |
Drill: justify why chunk boundaries fall where they do for lesson 159.
Practice
Practice List five adversarial prompts unique to your org’s nouns. — 159 Bump literals mindset by 39.