# The Model Has to Survive the Edge

_Published 2026-09-11._

An operator does not receive a model score. An operator receives a signal at a particular moment, from a particular camera, with a particular confidence, and has to decide whether it matters.

That difference is where many computer-vision demos stop being convincing. A notebook can show that a model recognizes smoke. A field system has to deal with a camera stall, dropped frames, inference latency, false positives, and an alert channel that may be unavailable.

## Training is not serving

Fire & Smoke Intelligence keeps the training environment separate from the runtime. GPU training with YOLO is one system. Exporting an artifact to ONNX or TensorRT and loading it on Jetson-class hardware is another. The export is a release boundary, not a file copied at the end of a notebook.

The runtime needs to know which model it loaded and whether the engine matches the deployment. If the artifact hash is wrong, refusing to start is safer than producing plausible-looking events with an unknown model.

## A detection needs a contract

A bounding box without time, camera identity, and confidence is not operational. The useful object is an event:

```text
frame -> preprocess -> inference -> detection -> event -> operator
```

The [Fire & Smoke case study](/work/fire-smoke-intelligence) keeps flight control outside this boundary. The system receives video and emits detection events; it does not pretend to control the aircraft or coordinate suppression logistics.

If inference overruns its budget, the runtime can reduce the sample rate while retaining the last good event. If the camera stops sending frames, a stream-health event tells the operator that the silence is a system condition, not an empty landscape.