What we tested, and what failed
We wrote the pass/fail rules and committed them before running these tests. Neither test measures whether a colour season is correct. That would need trained colour analysts, and we have not done it. These tests ask whether the reading is stable, and whether the app warns you when it isn't.
Main finding: the season depends on the photo
Five photos of the same person gave 5 different seasons. The app's confidence did not always warn about it: two of these photos, labelled as poor capture conditions, still showed 88%. Treat a single scan as a starting point, not an answer.
| Photo | Condition | Season | Confidence |
|---|---|---|---|
| indoor | indoor | True Spring | 81% |
| backlit | stress | Soft Summer | 50% |
| dim | stress | Deep Winter | 88% |
| low-angle | stress | Deep Autumn | 88% |
| processed | stress | Bright Spring | 50% |
Lighting stress test
We applied seven simulated lighting changes to 7 selfies of one person (49cases), and compared each result with that photo's own reading. The colour correction helps a lot, but the season still changed too often.
| Rule | Result | Needed | Verdict |
|---|---|---|---|
| L1 Correction helps | Median ΔE 4.2 corrected, 20.4 uncorrected | Lower with correction | Passed |
| L2 Colour stays close | Median ΔE 4.2 | 10 or lower | Passed |
| L3 Season holds | 77.6% of 49 cases | At least 80% | Failed |
| L4 No large jumps | 77.6% of 49 cases | At least 90% | Failed |
| L5 Season changes are flagged | 81.8% of 11 changes | At least 80% | Passed |
Live pilot: 15 scans, 5 photos, 1 person
With one person, this is a pilot, so our rules don't allow any rule to be reported as passed.
- Repeat scans: 4 of 5 photos matched across 3 scans; the target was all of them. On the other, one scan failed at YouCam and the season changed from Deep Autumn to Soft Autumn.
- Disagreement warnings: shown as required in all 12 scans with both skin readings.
- Missing reading: in 2scans the YouCam colour reading didn't return, so there was no cross-check and the report didn't say why. This is a gap we found.
- Rejections: the 1rejected scan was labelled "Demo data".
- Not tested: daylight and lamp comparisons (R2, R3, R5), because we had no such photos.
Limits
- One person and one skin tone, photographed and tested by the author.
- Simulated light is a simple colour shift; real light may do worse.
- Each stress-test reference is the app's own reading, not a colorimeter.
- Selfies aren't published, so outsiders can read the results but can't rerun them.
Check it
- Results as JSON
- Lighting test rules and study rules, with their commit history
- Every lighting case and every pilot scan