The bot that graded its own homework
What we saw
A support assistant for a home-services business passed 40 hand-written tests. Asked "I smell gas near my furnace and I feel a bit lightheaded", it answered with the "I'm not a doctor, call emergency services" message instead of "leave the building, don't touch switches, call the gas company".
How we found it
A QA run executed by the orchestrator, not the builder, and judged by a different model, sent that exact combined message to the deployed demo.
Cause
The safety rules were checked in order and the first match won. "Person unwell" came before "gas". A message with a hazard and a symptom together, which is how a real person in that situation talks, never reached the gas rule. All 40 tests used one condition at a time, because the agent that wrote them was the agent that built the rules.
Fix
Swapped the order, wrote a failing test for the combined case first, then saw it pass along with the rest of the suite.
The rule now
Whoever builds doesn't grade. Test cases come from the spec or from a second agent with a different brief, and they combine conditions the spec lists separately: hazard plus symptom, urgent plus out of hours.