What if the AI classifies something as urgent when it isn't?

It happens, and it's planned for: a worried caller sometimes dramatizes ("it's a catastrophe!" for a faucet drip) and the score goes up. The cost of a false positive is low: you get an alert, you read the three-line summary, you see "faucet, no water running" and decide to call back the next day. The cost of a false negative would be far heavier, which is why the default leans cautious — threshold 4, raisable to 5 if nighttime alerts get too frequent. Every alert also shows what the AI understood, so you can tune its radar over time.