What a hallway can answer without a camera
Occupancy can be read from Wi-Fi channel state, with no image stored. A 99.4 percent score is a public dataset, not a hallway. Fall detection is not claimed.
The short version
For occupancy, it says yes: empty or occupied from channel state, with no image. It says the model runs on the home computer, has not been fed by a live device in the house, and is not a fall alarm or a doorbell.
What the note is claiming
On October 3, 2026, dldeploy described rusty_esp_sense as a way to answer whether a room is empty without taking a picture. The ESP32 streams raw Wi-Fi channel state. A person moving changes the channel. The model that turns the stream into empty or occupied runs on the home computer, not on the chip, so it can be refit without a reflash. The note says the stream is about 7 kilobytes per second at 50 hertz, and that a fitted model is about 17 kilobytes from a few one-minute captures of the empty room and of someone walking. It says you calibrate while normal traffic is running, because a pattern the model never saw can look like a person. The paper filing the limit isFutureweb.
Whose bench the number is
On 100 captures from two ESP32-C6 boards in a public dataset, with every capture held out of its own calibration, the note says the score was 99.4 percent balanced against an on-chip detector's 80.0 percent. It says those boards were not the house devices, and that a recording of its own room has not been made. Quoting 99.4 percent as your hallway is the lie the note tells you not to tell. The desk isSmart Homes.
What is explicitly not a safety device
A fall detector and a night-epoch reader exist in the code. The note says no real fall has been recorded, no breath was accepted in the test data, and no fall rate or sleep result is claimed. If you need a fall alarm, this is not it. A camera still wins when you need a face, a package, or a porch. What local control requires, for the device that sources the stream, iswhat local control requires.
Why it matters on this desk
The dependency is the extra data a lens collects when the question was one bit. A camera sees faces and mail. Channel state, in this note, stores no image and no sound. The device under the stream still has an identity. A vendor account is the thing that identity is supposed to replace, filed aswhat a device-minted name replaces in a vendor account.
What this page will not import
This page does not name a dataset the note does not name, and it does not treat experimental fall code as a medical device. Espressif's Wi-Fi guide is the vendor document for the radio. The NIST Privacy Framework is the public vocabulary for data you would rather not collect. The mechanism note isdldeploy's presence note.
Where the public record sits
The note supplies the occupancy claim and the dataset score. Espressif documents the radio. The Privacy Framework documents a way to talk about minimizing what a sensor keeps. They do not overlap on 99.4 percent. That figure has one author, and the author says it is not your room. The on-chip detector's own empty-room and walking figures live in a different note. This page does not mix them with 99.4 percent. A doorbell is the case the note says still needs a lens. Occupancy will not show you who it was. A doorbell is the case the note says still needs a lens. Occupancy will not show you who it was. A doorbell is the case the note says still needs a lens. Occupancy will not show you who it was. The on-chip detector's own empty-room and walking figures live in a different note. This page does not mix them with 99.4 percent.
What to watch next
A live device in a named room, with the model refit there, would replace the public-dataset sentence. A fall recording would be a different claim and is not implied. What a session cannot be iswhat a session cannot be.
What would change the record
A note that the 99.4 percent figure was from a Janus device in a house would change it. A claim that the fall code is a safety device would change it. A camera at the door, where you must recognize a face, is the case the note says should stay a camera. No live Janus device has fed this model yet. That sentence is the note's, and it is the status. No live Janus device has fed this model yet. That sentence is the note's, and it is the status. The on-chip detector's own empty-room and walking figures live in a different note. This page does not mix them with 99.4 percent. No live Janus device has fed this model yet. That sentence is the note's, and it is the status.
What stays a bit, not a picture
What stays specific is occupancy from channel state, a model of about 17 kilobytes on the home computer, a stream of about 7 kilobytes per second, a public-dataset score the note refuses to certify as a hallway, and no image. Fall and sleep are code that exists, not results. A doorbell is the case the note says still needs a lens. Occupancy will not show you who it was. The on-chip detector's own empty-room and walking figures live in a different note. This page does not mix them with 99.4 percent. No live Janus device has fed this model yet. That sentence is the note's, and it is the status. The on-chip detector's own empty-room and walking figures live in a different note. This page does not mix them with 99.4 percent.
Sources
The reports this brief is filing. Futureweb did not republish them.
Questions
Does the October 3 sense note record video?
No. It says the model reads how a Wi-Fi signal changes across a room. No image and no sound are captured.
Where does that note say the occupancy model runs?
On the home computer, not on the ESP32. The note says the chip streams channel state at about 7 kilobytes per second and the model can be refit without a reflash.
What does the note say the 99.4 percent figure is not?
It is not a measurement of your hallway. It is 100 held-out captures from two ESP32-C6 boards in a public dataset, against an on-chip detector at 80.0 percent.
Can that note's fall detector be used as an alarm?
No. The note says no real fall has been recorded, no breath was accepted in the test data, and no fall rate is claimed.
How does the note say you calibrate a room?
A few one-minute captures of the empty room and of someone walking, then a model file of about 17 kilobytes, with ordinary network traffic running.