A city wants its permitting bot to stay concise and avoid sounding like legal counsel. Which optimization should the team apply?

A city permitting bot that must stay concise and avoid sounding like legal counsel should tune prompts and temperature so answers remain short and explicitly non-committal on legal advice. Prompt instructions set role boundaries (helpful permitting FAQ, not an attorney); raising temperature and top-p increases randomness and variability rather than concise, consistent guidance. Together they shape civic generative behavior toward safe, skim-friendly guidance with clear escalation to human staff for legal questions. Raising temperature and top-p so answers stay concise and non-committal increases randomness and off-the-cuff, counsel-sounding output—the opposite of the goal. Lowering frequency penalties so answers stay concise and non-committal encourages repetition and verbosity rather than concise guidance and does nothing to prevent counsel-sounding framing. Tuning logit bias settings so answers stay concise and non-committal only shifts likelihoods of specific tokens and does not enforce brevity or prevent legal-counsel framing. On the exam, prompt-plus-temperature tuning for concise, non-legal tone is the optimization; distractors include raising temperature and top-p, lowering frequency penalties, or tuning logit bias settings so answers stay concise and non-committal on legal advice. Caveat: even well-tuned bots can drift—add refusal templates and periodic evals for legal-advice-seeking prompts so regressions are caught after prompt edits. Operational check: run a suite of permit FAQ and “is this legal?” prompts, verify concise answers with non-advice disclaimers, measure length/temperature settings, and confirm system instructions still forbid legal counsel framing.