Evaluate LLM-based screening

[adelavega]

In #5 we will develop an LLM-based PRISMA-compliant screening workflow

We then need to evaluate its performance against neurometabench.

• First, evaluate the PubMed search itself, how many papers from the final inclusion list are even available? That is the maximum recall possible if only PubMed as a source
• Second, evaluate abstract only screening sep
• Third, evaluate final full text screening

Compute precision & recall for each step