Leniency bias
Faces that human raters scored 52 to 55 came back from the model as 62.
Technology
Getting a vision model to answer is easy. Making the answer true is the work. This is how we do it, including where we are still limited.
Five failures
We read hundreds of individual outputs, not just averages. These five failures turned up, and each was fixed in the system, not by asking the model nicely.
Faces that human raters scored 52 to 55 came back from the model as 62.
“T-zone oiliness” showed up on almost every scan: a prior pretending to be an observation.
Beard stubble read as blemishes, moles as dark spots, brow shadow as pigment.
One inflamed spot was counted as a blemish and again as a dark spot.
A person of about 65 was told fine lines were severe, 82, and overall skin health 35.
The mechanism
The order of the fields the model fills in is the mechanism. Generation runs left to right, so earlier fields constrain later ones: the model has to name what it excluded and count what it saw before it may score a single condition.
The score
The headline score is a deterministic function of the verified findings. Its weights were fitted by least squares on hand-scored production scans. The worst condition dominates, the rest add a smaller load, and each severity band has a ceiling: severe caps at 45, moderate at 72.
Examples: 62 becomes 52, 72 becomes 62, and 74 stays 74.
Weights come from human-scored examples, not from the model’s opinion.
The weights were fitted on a small set. A larger human-rated gold set is on our next-steps list.
The validator
After the model answers, a validator removes anything that is not justified, before a customer sees it.
An 82 stays an 82. From the 55 to 64 age band up, fine lines and dryness carry less weight in the headline and get the softer label “well established”. Age is a skippable dropdown and is never estimated from the face.
Why: age estimation is least accurate for older faces, and inferring demographics from a face carries legal risk in several regions. We chose the boring input over the impressive one.
Fairness
Stubble scored as blemishes hurt bearded users. It hurt one group most: razor-bump texture is most common in men of African ancestry. In aggregate accuracy it was invisible. We found it by reading individual outputs.
The fix is at the detection layer: confounders are named first, and any claim in that region is discarded.
Every regression test was verified to fail against the code that shipped the bug.
| Recommendation engine | 16 |
| Analysis validation | 23 |
| Scan key invariant | 14 |
| Explanation validation | 8 |
| Prompts load | 6 |
| Model response parsing | 14 |
| Trailing-object recovery | 10 |
| No em dash | 5 |
Honest limits
We would rather tell you than have you find out.
Scoring weights were fitted on a small set of scans. We are building a larger human-rated gold set.
Telling a person’s left side from their right is close to chance, so areas are reported in the image frame.
Telling blemishes from dark spots depends on the model telling a raised mark from a flat one.
Products are entered by hand today. Import from your store is not built yet. It is the next big step.
Because photos are never stored, a scan cannot be re-examined later. That is a deliberate cost of our privacy promise.
Detecting glare at capture time, before upload, is on the list.
The same photo can come back a few points apart. Some calls, like where flat pigment ends, are genuinely blurry. No setting removes that, so the headline score is computed by a fixed function.
We tell the analysis to judge each person against their own baseline, but an instruction is not a measurement. Checking accuracy across skin tones is part of the gold set work.
The portal runs on a tablet today. Automatic reset between customers, an idle timeout and a QR code to take the results home are not built yet.
Ready when you are
Book a demo and we will walk through a scan, the checks and a routine built from a sample catalog, then talk about how it fits your brand.