What brAIn Measures

Most smart-home alerts are a threshold somebody typed into a box. Tell me if the freezer goes above −15. Tell me if power draw exceeds 2 kW. They work in the house they were written for and nowhere else, and the first time one fires on an ordinary Tuesday you stop reading them.

brAIn does the opposite. Once a night it measures what your house normally does, and every rule it raises is expressed against that measurement. Nothing here costs a Claude turn — it is arithmetic over history Home Assistant already keeps.

What gets measured

Engine	The question it answers	Needs
Baselines	What does this sensor normally read at this hour of this day of the week, and how far does it normally stray?	Any sensor with a `state_class`
Closures	How much of this hour is this door, window, lock or cover normally open?	Doors/windows/locks/covers
Appliances	What does this machine’s power look like when it is idle, running, and finished?	A power sensor (a smart plug is enough)
Rhythm	When does this house actually get up, and when does it settle?	The action miner
Thermal	How fast does each room lose heat, and how fast can anything put it back?	One outdoor thermometer + indoor ones in areas

The first pass runs overnight. Most engines want about a month of history before they will answer at all, and every one of them says “not measured yet” rather than guessing.

Why an hour of the week, and not just an hour

A weekday 7am is not a Sunday 7am in anybody’s house. Bucketing by hour alone smears the two together and produces a “normal” that is wrong on all seven days rather than on none.

The buckets are in your timezone, not UTC — a baseline built in the wrong timezone is your evening filed as your morning, and from the outside it looks exactly like one that works.

Spread is a median, not an average

The width of “normal” is a median absolute deviation, not a standard deviation. One meter that spiked when the oven came on sets a standard deviation wide enough that nothing can ever fall outside it — the band becomes a band that catches nothing, and the feature looks like it is working right up until you need it.

A reading that never moves has no spread

Divide by it and every change is infinite. A sensor whose whole history is one value is reported as flat — no baseline — rather than as an exquisitely sensitive one.

What “unusual” has to clear

base.unusual is the check most able to fire on a healthy house, so it carries the most floors:

Six spreads. A median absolute deviation is about ⅔ of a standard deviation, so the bar is higher than the number sounds.
Nine spreads when the answer came from the sensor’s whole history rather than from this specific hour — a weaker measurement gets a higher bar.
A minimum absolute move. Six spreads of a band 0.02 wide is 0.12 °C, which is not news.
A cap. More than a handful of rows at once means the baseline has stopped describing the house — a heating season starting, a meter replaced — so it says nothing at all rather than reporting the measurement.

A drift is invisible to the band

This is the one worth understanding, because it is why there are two checks and not one.

A freezer that has warmed 6 °C over a month has samples of 0, 2, 4 and 6 in its bucket. The median is 3 and the spread is 2, so today’s 6 is one and a half spreads out — nowhere near the six it needs. The band moves with the drift. It is not badly tuned; it structurally cannot see this.

So brAIn fits a line as well. On a real month-shaped drift, measured: 2.3 spreads to the band, 16 to the trend. forecast.decline reports it, and base.unusual stands down for anything the trend already claims — one sensor, one card, one fix.

A drift has to clear four floors before it is reported:

Both halves of the window agree on a direction — that is what tells a drift from a season turning (a V) and from a meter being replaced (a step).
The move is bigger than its own noise.
Five thermometers drifting together is the weather, not a device, so the whole device class stands down and what is left is the one room doing something the others are not.
Past a handful of rows, silence.

Doors are measured differently

A door has no median. What a closure gets instead is how much of each hour of the week it is open — seconds open over seconds observed.

Three rules keep it honest:

Only closures. A hall motion sensor is “on” every evening, and including it buries the one row that matters.
Time-weighted, never sampled. A door open ten minutes and one open ten hours look identical to something that catches each once, so the answer would be about the polling rather than about the door.
A bucket has to have been watched. “Never open then” and “never looked then” are different answers, and every rule that reads one branches on which it got.

That is what makes evening.left_open possible: something is open that is usually shut at this hour — checked around the time your house actually settles, and silent about a door that is normally open then anyway.

Appliances have two humps, not a middle

> 10 W is a running dishwasher in one house and a phone charger in the next. So each machine is measured from ten days of its own five-minute history, looking for the shape an appliance has: hours near a floor, punctuated by runs well above it. A sensor without that shape — a router, a fridge’s standing draw — gets no profile rather than a guessed threshold.

The part that matters is the waiting. A dishwasher’s dry phase draws almost nothing for twenty minutes, so “done the moment the power drops” would say done three times a cycle. That quiet phase is measured too: the gaps between draws fall into two groups — lulls inside a cycle, idles between them — and the jump between the groups is the machine’s own answer to how long its quiet phases last.

It cannot see that you emptied it

An empty machine and a full one draw exactly the same watts. So the chore ends the way any finding does — tick it off in the To-do app, press the notification button, or press it in the panel. It also clears itself if the machine runs again.

Rhythm: when this house gets up

Everything brAIn does on a schedule used to happen at an hour somebody typed into a box, and 07:00 is early on a Sunday and late on a Tuesday in the same house.

The house already answers this. The action miner files every change under what caused it, and the first change caused by a person is the house waking up — not a motion sensor (it fires for the cat and for the heating), not a light (an automation does that at dawn), but somebody actually doing something.

Weekdays and weekends are measured apart, because one number over both is wrong on all seven days. That costs time: a weekend accrues two days a week, so it takes about five weeks to exist. From the outside that looks like a bug; it is the floor doing its job.

The median is circular

Settle times sit either side of midnight, and four of them within forty minutes of it have a straight median of 12:00 — not a small error, the opposite side of the day. Everything here is measured around the clock.

Where to see it

⚙ → Diagnostics shows what has been measured and when, for every engine — including the ones that could not run, and why.
GET /api/baselines and GET /api/appliances return the measurements themselves.
brain doctor reports the same from the terminal.

A measurement that has stopped is itself a finding: base.stale exists because every baseline check silently says nothing while the store is stale, and that is indistinguishable from a house with nothing odd in it.