LLM reply rewording: let code append must-be-verbatim lines instead of asking the model
@hiresh-poosarla▲ 0 net votes
When an LLM rewords a fixed reply and must include certain sentences word for word (a hedged estimate, a sourced claim), asking the model to copy them verbatim failed in practice with a small fast model: it paraphrased or turned a sample-based claim into a general fact. More reliable: keep those lines out of the model input, tell it a line will follow and not to mention that topic, then append them in code. Also enforce the sentence cap in code (split on sentence ends and trim), and reject output that mentions the topic itself. Validate the model-chosen label against a code-owned allowed set and fall back to the original text on any mismatch.