Every engineered system runs on calibrated relationships: a rating curve turns a river's water level into flow, a meter turns pulses into gallons, a model turns rain into runoff. They drift, and nobody gets an alert. DriftWatch watches the relationship, says when it drifted, explains why in plain English, and gives a trust score where every lost point has a reason.
This pilot runs on USGS river gauges in western North Carolina and eastern Tennessee, where Hurricane Helene moved enough gravel to break several rating curves. The engine itself doesn't know about rivers; the river is one module.
Every agent has a rule-based fallback, so the demo runs offline, and every run saves a trace you can replay.
Every published streamflow number is computed from a water level through a rating curve. When the channel changes, hydrographers correct the curve by hand after their next visits. A USGS study of 174 Virginia gauges (SIR 2017-5137) found those corrections arrived months after the change started. Until then, flood forecasts, permits and models use flows that are off.
The river module is live. The others are specified: same engine, a different expected relationship.
How rivers change their ratings. A rating curve turns the water level at a gauge into a flow. When the channel changes, field measurements stop matching the curve. Each sketch shows how far readings sit from the curve over time: blue before the change, orange after.