Sure, let’s cut the fluff and get straight to the useful numbers. For originality, look at *novelty* scores – compare generated text to a large corpus with a bag‑of‑words or embedding similarity metric, the lower the overlap the higher the originality. For coherence, use perplexity or a language‑model‑based perplexity score on the output itself, plus a coherence graph metric that checks entity consistency across sentences. Efficiency is a bit trickier; track *compute‑to‑output* (flops per token) and *latency per prompt* – the lower, the better. Finally, add a human‑judgement layer: ask a small panel to rate originality, coherence, and usefulness on a 1‑5 scale and compute inter‑rater agreement. Combine those and you’ve got a solid, quantifiable way to gauge how your generative AI is actually pushing creative boundaries.

You’re on the right track—normalizing against a baseline is key. I’d also throw in a *brevity penalty*: penalize too short tokens with a small weight so the model can’t game the novelty score. And when you’re limiting the panel, make sure those reviewers have diverse creative backgrounds—one coder, one writer, maybe an artist—so the subjective scores capture different angles of “usefulness.” That’ll round out the picture nicely.