What happens in those five seconds

The paw button sits in the middle of the nav bar, on every tab, and reads whichever dog is selected at the top of Home. Camera and microphone are asked for right there, on your first scan — never at startup.

The tail is the signal everyone watches and the one that explains least on its own. So the model is instructed to fuse face, body posture and vocal tone together, and told in the same breath never to read a dog from a single signal.

Then it shows its working. Every read carries a separate 0–1 sub-score for face, body and sound — how much each channel actually supported the verdict. The confidence beside it is the model’s own, and its instructions say in as many words that a low one is acceptable and trustworthy. It is a confidence, not a certainty, and there is no measured accuracy rate behind it — nobody has run the study that would produce one.

You come home, the bin is on its side, and there it is — head low, ears back, whale eye, the whole slinking performance. It reads as a confession. It isn’t one.

Dogs don’t feel guilt the way we do. That posture is an appeasement response to your tone and your body language: a dog reading the room, not a dog reviewing its afternoon. So the sixth mood is called Appeasing, and the correction is written into the model’s own instructions, in those words — dogs do NOT feel guilt; this is an appeasement response to the human’s tone and body language; frame it that way. The weekly report and Story Mode are handed the same line, so no part of Dogoos ever tells you your dog feels guilty.

Refusing is one half of getting it right. The other half is telling the model what it is looking at. A dog braced at the bathroom door and a dog braced at the front door are not having the same afternoon, and the footage can look almost identical. So before you record, say what just happened — one tap each, as many as apply:

Just had a treat · Said “walkies” · Dinner time · About to play · Vet / grooming · Bath time · New person / dog · Someone came home · Just woke up · Loud noise

They go into the prompt alongside the clip, and they stay attached to the read: the verdict card lists them straight back to you, and Story Mode picks them up when it writes your month. Mostly, though, they sharpen the read itself — which is the whole point of tapping them.

Eight everyday moments, each with the reaction to expect and a line on how to catch it. Record all eight and you have your dog’s own version of the eight situations that come round most — theirs, not a breed average, and sitting in your mood history alongside every other read. The app does not learn from them: a baseline you record today changes nothing about tomorrow’s read. It is a set for you to compare against.

In a two-dog house the obvious way to be wrong is to read the wrong dog. So your active dog’s profile photo travels with the clip as a labelled reference — downscaled to 512 pixels, fetched alongside the recording on a six-second leash, abandoned without complaint if it’s slow or missing. A photo is never allowed to be the reason a read fails. It goes only when personalised insights are on, because it is identifying data; turn that off and the AI gets the clip and nothing else — no photo, no profile, not even your context tags.

The model also hands back a timestamp: the strongest single frame in the clip, chosen under rules about what doesn’t count — rear legs only, the dog leaving frame, motion blur, occlusion. That frame becomes the keepsake in your album, shown uncropped on purpose, because re-cropping a fast-moving dog is how a good full-body moment turns into a misleading partial one. When the clip won’t give that exact frame up, the app takes a plain camera still instead rather than leaving you an empty album slot. The video itself doesn’t stick around: Dogoos does not store the clip as a video.