Another use case for Jev. This time: 911 dispatch. After Jevton, I wanted to see how far a decision model that never writes a word could go on real data. So I put Jev on live emergency calls, with real dispatchers as the answer key and a dumb baseline on screen so it can't hide. Jev Dispatch pulls 911 calls from Detroit Police and San Francisco Fire & EMS, straight from the cities' open data. Jev sees only what the calltaker logged: "SHOTS FIRED", "SUICIDE THREAT", "HANGUP CALLS", the neighbourhood, the time. Then it decides: priority 1 to 5, is a life at risk, how soon will police arrive. The dispatcher's real priority and the real response time are the answer key. Next to every Jev score there's a second column: what you'd get by always guessing the most common answer. Green if Jev beats it. Amber if it doesn't. No cherry-picking possible. The numbers: 3 decisions a call, 10 calls a request, about 300 ms back, a few cents for a whole day of a city's emergencies. Zero text generated. Probabilities in, decisions out. What I learned building it. In San Francisco 79% of calls are just "Medical Incident at an intersection". No caller words, no symptoms. Jev lost to the baseline there, and it should have. So SF only gets forecasting questions now: which unit goes first, how many, how soon. Detroit's calltaker codes carry real signal, so Detroit gets the full triage. Garbage in, garbage out applies to decision models too. The fix isn't a bigger model. It's giving the small one what the human had. Watch Jev triage a city in real time: dispatch.chizi.app