A clerk of courts needs fee quotes that never invent amounts outside the statutory fee table. Which orchestration approach should the developer implement?

A clerk of courts that needs fee quotes that never invent amounts outside the statutory fee table should combine LLM drafting with a rules engine that enforces those tables. Friendly explanation language can be generative, but dollar amounts are statute-bound. The rules engine (or equivalent deterministic fee service) is authoritative for amounts; the LLM explains process and context around those stamped values. That hybrid orchestration prevents creative fee invention while still giving residents clear, readable quotes.

Grounding LLM drafting with retrieval-augmented generation over fee schedules still lets the LLM compose the final dollar figure from retrieved passages, and a generative model can paraphrase, misread, or extrapolate a number from retrieved text even when the source passage is correct—RAG improves grounding, but it does not guarantee the model never invents an amount the way a deterministic rules engine does. Fine-tuning the LLM on past clerk invoices to align fee quote calculations trains the model to produce outputs that statistically resemble historical fees, but a fine-tuned model is still a probabilistic generator that can extrapolate a plausible-sounding amount outside the current statutory table, especially after the table changes. Adding a second LLM review pass to verify fee quote amounts before court filing may catch some obvious errors, but a second model reviewing another model’s output is still generative-on-generative; it can misread the schedule, approve a plausible-but-wrong amount, or hallucinate confirmation, and neither pass is the deterministic authority the stem requires over the actual statutory numbers.

Exam caveat: for statutory fees, determinism owns the numbers and the LLM owns the prose—never invert that authority, and don’t mistake grounding with RAG, fine-tuning for calculations, or a review pass to verify amounts for the deterministic guarantee only a rules engine provides. Operational check: bind fee quote generation to the statutory table/rules engine, reject or escalate when a fee code has no row, display only engine-stamped amounts in resident quotes, and audit sample quotes against the published fee schedule after each release.