Time: 30–45 min
Goal
Feel the harness as a seatbelt: change behavior → suite fails → you decide keep/revert.
Drills (pick ≥2)
- Baseline bug: cancel even when status is
shipped→ policy goldens must fail - Missing refuse: unknown order invents a cancellation → refuse golden fails
- Double cancel: second cancel increments a counter / errors loudly → idempotent golden fails
- Tool skip: answer “done!” without calling
cancel_order→ tool-expect fails
For each: run suite before/after; write one line in your study log.
Reflection prompts
- Which failure mode would users hit first in a real support agent?
- Would a typical jam “verify green” have caught this? (usually no — different lane)
Log: did: Day 4 broke X caught by G0xx · takeaway = favorite catch.