It's tempting to point a language model at a competitor's activity and ask it to rate the impact. It's also a mistake — because the moment your score comes from a model, it stops being reproducible, and a number you can't reproduce or explain is just an opinion wearing a decimal point.
Determinism is trust
The IDØ impact score is computed from the diff and the measured public reaction — the same inputs, the same number, every time. You can open the features that produced it. That's what makes it something you can build a decision on instead of something you have to take on faith.
Where the model does belong
The model appears exactly once, at the very end, to phrase the brief in plain English. It writes prose; it never invents the score. Generation is a fine job for a language model. Being the source of truth is not.