Skip to content

2026.10.2

Latest

Choose a tag to compare

@Hankanman Hankanman released this 06 Oct 12:39
· 21 commits to main since this release
f9d942f

A big thank you to those who have chosen to sponsor my work!
You can sponsor me here

This release makes what Area Occupancy learns more trustworthy, and starts gathering the evidence for the next big step. There's no database rebuild.

🐛 Fixes

Rooms only learn from time they could actually see (#574)

Until now, every hour counted towards a room's learned habits, including hours when its motion sensor was offline (a flat battery, a Zigbee dropout, a sensor not yet paired). Those hours were learned as "empty", which pulled the room's learned likelihood of being occupied down after any outage. A TV reporting "off" also counted as watching the room, though it can't see that a room is empty.

Now:

  • Observed time only: a room learns only from time when one of its motion sensors was reporting, plus any time it was occupied. Short gaps (under 10 minutes, like a restart) are ignored.
  • Floors and All Areas count a moment only while every room in them was observed.
  • Away time: if you've set an Away mode entity (Configure → Global Settings), time while it's on is left out too, so a holiday doesn't teach your rooms that they're usually empty.

Rooms whose motion sensors report continuously won't notice a difference. Rooms that had outages will learn a more accurate likelihood of being occupied.

A one-time repair also clips the last few overlapping sensor records left over from the 2026.9.4 sync issue.

🧪 Shadow mode: proving the maths before it changes your occupancy

Four features are now recording what they would do, without changing any reading. The goal is to prove each one against real homes before it's allowed to affect occupancy. You can see what they're recording in the diagnostics download (Settings → Devices & services → Area Occupancy Detection → ⋮ → Download diagnostics), under each area's current block.

Feature What it would do What it's recording
Accuracy and suggested threshold (#499) Suggest, and later set, each room's occupancy threshold from how well it has been tracking reality How often each room's decision matched what its motion sensors confirmed, its false on/off rates, and the threshold the data argues for. A daily history is now kept across restarts.
Continuously learned priors (#500) Learn each room's habits as they happen, instead of from an hourly database recalculation The live estimate beside the database one, and how many days they've agreed. It now restarts itself if a room's sensors change, so stale history can't skew it.
Learned sensor weights (#501) Learn how much to trust each sensor in your home Each sensor's learned weight next to its current one, with its own sample count. Removed sensors are now forgotten.
Away mode lowers occupancy (#584) While your away mode entity is on, drop every room's baseline to 1%, so a TV on a schedule, a drifting sensor or a robot vacuum can't hold an empty house occupied. Real activity still shows: one motion sensor still reads about 99%. How long you were away, how often each room was really occupied while you were away, and above all how long away mode would have read a room as empty while someone was actually there. That number must stay near zero.

When will they go live? The plan is for these to start affecting occupancy in 2026.11.1, once there is more than a month of data from a real install proving the maths holds up. Each one goes live only if its own evidence clears its bar. If the data says one isn't ready, it stays in shadow mode, and the release notes will say why. Nothing changes for you in the meantime.

Want to help? Away mode can only gather evidence if you set an Away mode entity (Configure → Global Settings): an input_boolean, binary_sensor or switch that's on while everyone is out, such as a "nobody home" template from your zone.home, or a vacation toggle. Don't base it on Area Occupancy's own sensors. Diagnostics exports from homes running these are very welcome on the issues above.


Upgrade check: scripts/harness upgrade from a lived-in 2026.8.1 base to this release reported 0 red flags and all 11 checks passed.

Full Changelog: 2026.10.1...2026.10.2