# How a case is made

> Our expert network writes and reviews every case. Then we find where leading models fall short on it. Each gap shows exactly what a model still needs to learn, and that is what makes a case valuable.

## 1. Write

A practicing dentist writes a clinical question where the outcome depends on patient context and expert knowledge. Often this involves a health factor, a subtle choice between comparable options, or a pitfall that may be unsafe. The same dentist writes a gold-standard answer and a rubric that scores any answer against it.

## 2. Test

We give the case to leading models, with set criteria for model difficulty. This could be a pass rate, a minimum score, or another measure of success our customers choose. Each model solves the case several times, to show that its knowledge gap is real. These solutions are the model attempts.

## 3. Grade

Our rubric defines the standard set by our expert network. Each item lists what a good answer should say, with penalties for unsafe or evasive replies. An LLM-as-a-judge acts as the grader and applies that standard across all the model attempts. No model grades its own answer. The result is a score for how well each model answered.

## 4. Keep

When a score is below a set threshold, we deem the case difficult. Our rubric also exposes specific gaps, such as treatment blind spots or safety pitfalls, which point to a valuable direction for training. Those gaps are what our customers use our product for: to measure and improve their models.

See every step on the sample case: https://frontierdental.io/case-study.md