Most of the lights in the house are colour bulbs that follow a slow curve through the day, warm and dim first thing, bright and neutral in the middle, warm again in the evening. The stairs are not. The stairs are a plain dimmable LED strip on a real dimmer, no colour, no cleverness. And for weeks they looked wrong at every hour, in a way I could see but could not put a number on.
The stairs were the thing I was least sure about when I bought this place, so of course they are the thing I have spent the most time on. Two flights inside one home still strikes me as a design decision somebody made for a reason I have not worked out. They are also the only part of the house where getting the lighting wrong has a consequence beyond mild annoyance, which is presumably why I keep circling back to them instead of doing the hallway.
The problem is that fifty percent on a dimmer and fifty percent on a smart bulb have nothing to do with each other. They are two different manufacturers' ideas of a middle. So I stopped guessing and measured it.
The rig
I borrowed a light sensor from another room, put it in the staircase, and used a second one near the kitchen downlights as the reference. Then I wrote a script to sweep both fixtures through their range, wait twenty seconds at each step for the output to settle and the sensor to report, and record what it saw. It ran unattended overnight and took about forty minutes.
It fell over twice before it completed. The bridge will only hold a few simultaneous secure connections, and I already had two long-lived event streams open, so a script that opened a fresh connection for every single request ran it out of capacity and everything stalled. Once I rewrote it to hold one connection and take turns on it, the problem vanished entirely and has not come back anywhere else in the system.
What the numbers said
Below about twenty five percent, the stair strip produces essentially nothing. Around one lux. Not dim, not a soft glow, just off with extra steps. Everything I had been asking for in the small hours was in that band.
It peaks at eighty five percent. Above that it gets darker again. A hundred percent measured about four hundred and fifty seven lux against five hundred and two at eighty five, which I assume is the driver heating up and pulling back. I have checked it twice because I did not believe it either.
I am aware that I spent an evening building a measurement rig for a staircase in a home I have lived in for a few months, and that nobody who visits will ever notice the difference between the version of this I shipped and the version I could have got by nudging a slider until it looked about right. Somebody else in the house walked past while the sweep was running, watched the strip step up one percent at a time in silence, and asked if it was broken. That is roughly the correct reaction. I did it anyway because the not-knowing was bothering me more than the light was.
The output for that is a lookup table, so when the rest of the house is at a given point on its curve the stairs get whatever percentage actually produces matching light. Fifty on the bulbs is forty six on the stairs. Anything above seventy five is capped at eighty five, because asking for more makes it worse.
The bit the measurements did not catch
I set the overnight level to twenty one percent, which the table said was a sensible floor, and it worked when I tested it. Then a few nights later I watched it light up properly and fade away to nothing over about five seconds.
These drivers need one amount of power to strike and a different, lower amount to keep going, and both thresholds are higher than the table implied. They will not ignite below about fifty percent at all, and they will not hold below about twenty seven. Twenty one lit, then died, every time, which is exactly what you would expect and nothing I measured would have told me.
So now every motion trigger for the stairs strikes at fifty percent, holds it for just over a second, then settles to whatever the real target was. You cannot see it unless you are looking for it. It is not the solution I wanted, and I spent a while trying to find a cleverer one before accepting that the hardware simply behaves this way.