At 03:11 on a Tuesday, hard bounces on our marketing vendor tripled in ninety seconds. Nobody was awake. Nobody needed to be.
The breaker tripped at 03:12, traffic weighted onto the EU fallback, and the queue drained without a single lost message. The LLM diagnosis — a reputation dip tied to one campaign segment — was waiting in the incident channel.
The postmortem wrote itself from the routing ledger. Total human involvement: one approval tap, from a phone, at 03:14.
The timeline, reconstructed message by message: 03:11 the storm begins. 03:12 the breaker opens. 03:12:47 the last in-flight message leaves the degraded vendor for the EU fallback — those are the 47 seconds of the title. At 03:14 a phone buzzes in a dark apartment; one tap approves the diagnosis and files the rule change. By 03:20 the bounce rate is back inside the normal band, and the only artifact by morning is a postmortem nobody had to write by hand.
We route 1.8M emails a month and hold a 99.2% delivery rate. Incidents like this are how the number survives contact with reality: the machines handle the seconds, humans handle the judgment. Across the six incidents in this lane, that division of labor never once needed a meeting to activate.