Outlines

Why "don't say X" prompts backfire, and how a tiny grammar-constrained model can scrub a script without breaking its grammar.

A 27B model promises answers without the preamble. Its benchmark is homegrown — and the behavior it targets has a paper trail.