15–16 October 2026
One night, unedited
Every photograph this feeder took between dusk and the following morning. Not a selection — all fourteen, including the eight that turned out to be nothing and the two animals it could not name.
Recorded on the build of 15 October 2026. Where something here has since changed, the faults below say so and give the date — an old night left standing without that is a report of a product that no longer exists.
The whole night
Warmer than its background means brighter here. The six marked frames are the two real visits; the rest is what an ordinary night looks like.
First visit — 20:12, four seconds
Something crossed, warm and low
It was followed for ten seconds and travelled most of the frame. The contrast climbed from +4.9 °C above the wood to +14.6 as it came nearer — and the lumpiness rose with it, from 1.8 °C of variation across the shape to 10.5. Warm ground does not do that; a body does.
Not named. At its largest it covered nine tenths of one per cent of the frame. Nothing that small is offered to the classifier — below about a twentieth of the picture it answers “blank” whatever is in front of it.
Second visit — 21:56, forty-five seconds
And something walked towards the camera
Forty-five seconds, three hundred pixels of travel, and a shape that grew from 125 pixels to nearly four thousand. The contrast rose the whole way:
+5.1 → +8.9 → +10.4 → +12.2 → +15.9 → +18.1 °C
It finished at 36.0 °C against an 18 °C wood, close enough to fill a good part of the frame: hunched, snout down, with a textured back.
Still not named. This one was large enough to ask about, and the classifier was asked three times. It answered blank at 0.999, then blank again. Its best guess at any point was “greater hedgehog tenrec” — a Madagascan animal — at a confidence of 0.04. A model trained on daylight photographs, shown heat, is out of its depth and says so.
Honestly
What this night got wrong
A good night by any fair reading, and still three things we would rather it had done better.
- Thirteen notifications for two animals. Each frame of a visit
alerted separately. One visit should produce one notification that improves as
more is learned, and this did not.
Fixed 21 October 2026. A visit now gathers its photographs: later ones carry the same notification id, so the phone replaces the message in place and only the first one makes a sound. Running at the feeder since; not yet proved over a whole night here, because that needs another night like this one. - Neither animal was named. Heat found both, measured both and
photographed both. The species model had nothing useful to say about either.
Still true, and not a bug. At its largest the second animal covered nine tenths of one per cent of the frame. Nothing that small is named — see By heat for where the size limit falls. A closer animal is named from a thermal frame; these two never came close enough. - One animal was recorded as three visits. The second changed
identity twice while it walked, because a warm body breaks into pieces for a frame
or two and each piece looks briefly like a new arrival.
Open. Nothing has changed here as of 23 October 2026.
Why publish this
Because the alternative is a photograph of a stag at golden hour and a number nobody measured.
An ordinary night is mostly nothing. Twenty-five warm shapes were followed here and twenty-three were vegetation, cooling ground, or a body too far off to be sure about. Knowing that in advance is more useful than a highlight reel — and if we are willing to show you a night where it named no animal at all, you can believe us about the nights where it works.