Black-box AI vs deterministic scoring: why it matters

There are two kinds of facial analysis tools on the market. One takes your photo, runs it through a vision model, and prints a score. The other takes your photo, extracts measurable parameters, runs threshold math against published clinical bands, and prints a score with the math attached.

They look identical from the outside. They are not the same product.

What black-box scoring does well

Modern vision models are accurate, fast, and getting cheaper every quarter. For consumer applications where the user does not interrogate the output (filters, makeup overlays, social rankings), a black-box model is the right tool. The user wants a number, the model returns a number, both parties are happy.

Where it fails

The failure modes are all variations of one problem: the output cannot be defended.

• Training bias. A model trained predominantly on Caucasian faces will systematically misjudge faces of other ethnicities. The user has no way to see this in the output.
• Reproducibility. Run the same model on the same photo twice, and you may get slightly different scores. Run it after a quarterly model update, and you may get noticeably different scores.
• Defensibility. A user asks "why is my brow score 6.1?" The answer is "the model said so." That answer does not survive a conversation with a dermatologist.
• Edge cases. Atypical faces (post-procedure, partial paralysis, asymmetric scarring) confuse models in unpredictable ways. The model will still return a confident number.

What deterministic scoring looks like

A deterministic scorer never lets a model output a score. Instead it executes a fixed sequence:

1. Detect landmarks. Reproducible on the same photo every time.
2. Compute parameters from coordinates. Pure math, no model in the loop.
3. Compare each parameter against a threshold band drawn from a peer-reviewed paper, keyed to the user's gender and ethnicity.
4. Map the comparison to a numerical band score and a categorical tier (Elite, Great, Good, Fair, Needs Work).
5. Aggregate module scores using a fixed weighted formula.

Every step in this sequence is reproducible, auditable, and citable. If a user asks "why is my brow score 6.1?", the answer is: peak rise was measured at 4.8mm against the gender-typical Carruthers band of 5.0 to 7.0mm, which placed the parameter in the 38th percentile, which mapped to a band score of 6.1 under the Eyebrows module formula.

Where Claude appears in Facet

Claude is invoked only after every score is computed deterministically. Claude writes the qualitative observations ("the right brow peaks 0.7mm higher than the left, contributing to a small asymmetry that reads as expressive rather than imbalanced") and the natural and clinical protocols. Claude never decides scores. A function called stripClaudeScores actively removes any score field Claude tries to return.

The dividing line is hard. Scores are math. Words are language. Two different jobs, two different tools.