Everything runs on this server. No text leaves the machine and there are no per-scan costs.
Fast mode measures stylistic structure: sentence-length variation, paragraph uniformity, connective scaffolding, boilerplate phrasing, hedging, repeated openers, concrete detail.
Deep mode adds what a language model finds surprising — perplexity, how often the next word was its top choice, and how much that varies across sentences. In theory that is harder to edit around than style, but it has not been validated here yet, so it is off by default and marked experimental.