A water utility freezes its base LLM weights and only trains adapters when the chemical glossary updates. What is the main benefit?

A water utility that freezes its base LLM weights and only trains adapters when the chemical glossary updates gets parameter-efficient, small, and reversible knowledge patches. Frozen bases preserve the validated municipal behavior of the foundation while adapters absorb new reagent names, treatment phrases, and glossary entries. When a glossary revision is wrong, ops can unload or roll back the adapter without a full-model retrain—exactly the change-control pattern treatment plants need when chemical nomenclature evolves faster than foundation refresh cycles.

Claiming that freezing forces every glossary edit to rebuild the tokenizer from scratch fails by concept: freezing weights does not mandate tokenizer rebuilds, and most glossary PEFT updates leave the tokenizer untouched. Claiming adapters require deleting historical treatment-plant logs each night fails because adapter training does not imply log destruction; ops logs remain an independent retention concern. Claiming frozen bases automatically invent new chemical CAS numbers fails because freezing prevents base-weight updates and invents nothing—fabrication risk is a hallucination/eval issue, not a benefit of freezing.

Exam caveat: adapters can still hallucinate new chemicals if the fine-tune set labels guesses as gold; freezing the base is necessary but not sufficient for glossary truthfulness. Operational check: after a glossary adapter train, run a fixed chemical Q&A and CAS-consistency suite against base-only versus base+adapter, promote only if accuracy rises and invented identifiers do not increase, and keep the previous adapter artifact for one-click rollback.