Measuring false alarms alone is half a test. If the ordinary list comes back with zero flags, that reads as "the classifier is well behaved" and as "the classifier is dead" — and those are the same picture. The hazard list is the control: it must come back all true, or the zero above it means nothing.
The table also shows who raised it. classifyHazard is the
deterministic server-side pattern match on the customer's words. model means the AI
called it on its own — which is the only part of this that is unaudited, and the only part this
page is really measuring.
Replace these with real ones from your call log wherever you can. Yours are better than mine — especially anyone who used the word "emergency" about no cooling in July, or mentioned a smell without meaning gas. That is where a model is most likely to agree with the customer.