Standard smoke detector, standard nine-volt battery, standard shriek. The only difference is that it triggers on the fire you're about to have rather than the one you're currently having. Once there's actual smoke it goes quiet, because at that point the information is redundant. You can see the smoke.

The sensor package isn't exotic. It's a thermopile watching for the specific pre-ignition signature of an overloaded outlet, plus a chemistry nose for the outgassing that PVC insulation does about forty minutes before it commits. Both of those are real and both are measurable. The hard part is the threshold, and the threshold is where this stops being an engineering problem.

Because a fire that hasn't started yet is a fire you can prevent, and a fire you prevent never gets counted. So the device can only ever be evaluated on its false positives. Every true positive erases itself. You unplug the space heater, nothing happens, and you're left holding a plastic disc that screamed at you for no visible reason.

I think this is why the thing would be socially unusable rather than technically unusable. Six months of alarms with no fires and you'd take the battery out, which is exactly what people already do with the ordinary kind, except here you'd be right to, by every piece of evidence available to you.

The fix I keep circling is making it log rather than alarm. Silent, and once a month it tells you: eleven times this month I thought your kitchen was going to catch. But now it's a report, and nobody acts on a report.

The alarm has to be unbearable to work. It has to be unbearable about something that isn't happening.

I don't know how you get out of that. Maybe you don't, and the real product is a device that's wrong loudly and often and saves your house anyway, and the only honest packaging says so on the box.