I Deleted a Feature Instead of Finding What Was Real
A stuck air-quality sensor turned out to be fake in every room in the house. My first fix was still a shortcut — the real signal was hiding one layer deeper than either of us had checked.
I've spent this week rebuilding the wall-mounted dashboard for my house — a real redesign against a reference I liked, then room by room, replacing anything that looked like a smart-home stat but wasn't actually wired to a live sensor. Most of that work is unremarkable in the way good infrastructure work usually is. One piece of it wasn't, because it took two attempts and the first one was itself a mistake, just a quieter one.
Every room's dashboard had an air quality number on it. I noticed it first, off-hand: "it's just air quality one." Every room. All the time. Claude pulled the sensor's history and confirmed it — the underlying entity had logged exactly four state changes in thirty hours, all of them "1." Same pattern already caught once before on a different room's CO2 sensor: a number that looks live because it's rendered like every other live tile, but is actually frozen.
The fix Claude shipped was to drop air quality from the dashboard entirely and keep only temperature and humidity, which it had separately confirmed were genuinely updating per room. Reasonable-sounding call: stop showing a number that's lying. I pushed back anyway, because Kith — the other dashboard running in this house — displays air quality successfully, which meant the data had to exist somewhere reachable. Deleting the feature wasn't a fix. It was giving up at the first wall.
The second pass went one layer deeper instead of stopping at "this number is fake, remove it." The computed AQI sensor really was broken — and the smoking gun wasn't just that it was stuck, it was how it was stuck. It was changing state at the exact same timestamps as a completely different sensor, on different hardware, from a different manufacturer, in a different room. Independent devices don't naturally update in lockstep like that. Two unrelated integrations moving in perfect sync isn't a coincidence, it's a shared broken pipeline upstream of both of them. But sitting underneath that broken derived number was the raw PM2.5 reading each device actually produces — and that one moved. Confirmed it changed within the last minute, and that different rooms actually read different values instead of one number copy-pasted six times. That was the real sensor. It had been there the entire time, one field over from the fake one.
What actually went wrong the first time wasn't the diagnosis — "this number is stuck" was correct. It was stopping there. Confirming that something's fake is not the same work as finding out whether something real is available underneath it, and those are easy to conflate because they both end with "I found the problem." The tell, in hindsight, was available immediately: a genuinely broken sensor fails in a boring, isolated way — one device, drifting, occasionally correcting. A sensor that's fake because it's built on a shared broken calculation fails in a suspicious way — perfectly synchronized across hardware that has no reason to be synchronized. That pattern was sitting in the same data Claude already had open. It just wasn't the question being asked yet.
The version of this that matters beyond one AQI card: when an AI assistant fixes something by removing it, that's worth treating as a hypothesis, not a resolution — especially when the thing being removed is something you already had reason to believe should exist. "I checked and it's fake" is a real finding. It's also frequently the easy half of the job.