Adjectives do not transfer, and a few examples do not either
Before building anything, we researched how style replication actually works, rather than guessing at it. A few findings decided the shape of the system. A 2025 study testing frontier models imitating more than four hundred real authors across news, email, forums and blogs found that in-context examples alone still drift to a generic, averaged tone, easy to catch as AI, especially in informal writing, and that more examples past four or five barely help. Separately, the standard method in authorship analysis (comparing the frequency of the roughly hundred and fifty most common words: and, so, just, like, I, the) shows that identity lives largely in those small words, independent of topic, which means a voice can be measured and a draft can be scored against it. A retrieval benchmark showed that pulling a person's own relevant past writing for a specific task beats generic personalization, and picking the right few examples matters more than adding more. And a study on learning a person's editing preferences found that turning their edits into short, retrievable rules beats both raw edits and fine-tuning on cost and accuracy.
Those four findings ruled out the two easy paths: no fine-tuning, and no adjective list standing in for evidence.
Local, layered, and measured against a real baseline
The system that came out of that research runs entirely on the person's own machine, reading their own texts, email, dictation and recordings, and produces five things, none of which are a fine-tuned model. A counted fingerprint per register (texts read differently from email, which reads differently from dictation), covering filler and buzzword rates, sentence length, punctuation and openers, each measured against an AI baseline built from a model's own replies rather than an assumption. An exemplar bank of real passages, tagged by register, recipient type, intent and era, retrieved three to five at a time to match the task at hand rather than pasted wholesale, because the research showed no gain past about five. A contrastive rule layer built from the hand edits a person actually made to a machine's cleaned-up drafts, in the shape of short, retrievable corrections rather than a giant style essay. A mind profile of principles, triggers and how the person treats people, every claim tied to a count or a quote from their own corpus, checked by an adversarial pass that tries to break every claim. And a scorer that measures stylistic distance from a draft to the person's own register versus the AI baseline, used to check drafts, never to write them.
How we came at this one
The question asked first was what actually carries a person's identity in their writing, independent of what they happen to be writing about. That question fit because the goal was never to describe a person's style in words, it was to measure it, and only a measured thing can be checked against a real draft later.