CLIP and CLIPScore: One Model, Two Jobs in the Same Pipeline

A new engineer on your team asks: "Why does our data pipeline need CLIP if we're building a diffusion model, not a classifier? And is the CLIPScore we report in our eval dashboard the same CLIP, or a different model?"

1. Explain what CLIP is and how it's trained, at the level you'd use in an interview.
2. Answer both of the engineer's questions directly.
3. Propose a concrete scenario where CLIP-based caption filtering in data prep and CLIPScore-based evaluation could disagree with each other, and explain what that disagreement would tell you.