How do you choose between rules, a trained classifier, and an LLM?

Run all three on the same held-out fragments before anyone picks the model, the way Lesson 09 of Becoming Forward Deployed Engineer does: in its sample run on Meridian's intake documents, rules scored 97.5% with three no-answers, a TF-IDF classifier 100%, and a 3B local model 87.5% at about 5,805 times the latency of the rules, so the model came third. Then add two columns the table does not have. A slice column, cut by whatever the training data is skewed on, because a classifier of mine scored 89% on held-out data and 61% in production within six weeks when an English-heavy training set scored non-English customers as angry, and a feature of mine passed on average across 200 real cases and failed on one slice that was a fifth of traffic. And a date: re-run the whole table on production rows at week six, by slice, with the no-answer count on the dashboard, because it moves before the accuracy does. Route rules first and escalate abstentions; the 100% row with zero no-answers is the one to fear, since it cannot say it has never seen this before.