Sixteen look directions¶
The v2 pet contract adds two atlas rows of look frames: sixteen still poses in 22.5° clockwise increments starting from straight up (000). The host picks the nearest one to the pointer's angle, which is what makes a v2 pet feel present instead of decorative.
How we know they're right¶
Look frames are easy to get subtly wrong (a pupil that drifts the wrong way at 292.5° reads as "shifty" rather than "up-left"), so they were tested three ways, and the results are checked in under qa/:
| Test | What it checks | Result |
|---|---|---|
Semantic review (direction-semantics.json) | An independent visual pass judging each of the 16 frames against its expected direction | 13 pass, 3 warnings (112.5, 157.5, 292.5: correct axis, subtle vertical component), 0 failures |
Blind pair validation (direction-blind-validation.json) | Frames shuffled into anonymous A/B pairs; a classifier with no labels decides left/right and up/down | 11 of 14 pairs pass; the hard-gated cardinal pairs (090/270, 000/180) all pass; 6 warnings on ambiguous diagonals flagged for human review |
Continuity (look-continuity.json) | Pixel diff and centroid shift between each adjacent pair around the circle | Median 2035.5 px changed between neighbours; the largest step (337.5 → 000, 5421 px, 8.2 px centre shift) is the loop seam and was accepted after review |
Those warnings are real and left in place on purpose. The frames read correctly in motion; the classifier's uncertainty on the shallow diagonals is what you'd expect from a character whose "up" is mostly eyebrow. If you can make 292.5° read better without breaking its neighbours, that's a welcome PR.
The whole story with numbers is on the QA evidence page.
For the ChatGPT web build¶
Web upload uses the v1 layout, which has no look rows, so the web pet has all nine animations but doesn't track the pointer. See Pick it in ChatGPT & Codex.