Architecture
Mimesis MCP is a local-first Model Context Protocol server that runs entirely offline. It integrates with Claude Desktop and the Claude Code CLI, giving a language model tools to read an author’s stylistic corpus and to have its own drafts vetted against it. Local ONNX embeddings and SQLite FTS5 indices mean no private manuscript ever leaves the machine.
Manuscripts are ingested from PDF or DOCX, cleaned of boilerplate, and packed into overlapping ~200-word chunks. Passages are embedded with a CPU-optimised BAAI/bge-small-en-v1.5 ONNX model and stored alongside an FTS5 table, and retrieval fuses cosine similarity with BM25 keyword search through reciprocal rank fusion.
The surface scrubber, and its ceiling
The first version of this system did what most style tools do. It filtered a banlist of overused AI diction, profiled hedging density, measured sentence-length burstiness to catch the flat uniform rhythm characteristic of model output, and banned em-dashes outright.
That layer is necessary and it is not sufficient, and the interesting part of this project is the evidence for the second half of that sentence.
Three measurements made the case. Generated text scored AUROC 0.691 on the surface features a word-level scrubber can see, against 0.884 on cadence features it cannot — the tells being removed were already nearly matched, and what separated the prose was invisible to the filter. Optimising toward the author’s corpus mean produced text scoring AUROC 0.316, below chance, meaning generations sat closer to the author’s centre than the author’s own writing does. And when the stylometric fingerprint was finally validated against a human ranking, it agreed 60% of the time, identical to a control that just counted words.
A blocklist can remove the cheap tells. It cannot make prose human, because the tell that survives is the shape of the sentence, not its vocabulary.
Rhetorical detectors
The current system adds a layer that reads sentence structure rather than word choice. Each detector was added after a specific draft passed every existing check and was still rejected on sight in blind review.
- Self-explanation. A clause arrives at a point, then instead of stopping
appends a clause narrating what the point was — the
, which meanshinge, or a reverse pseudo-cleft with a nominalised subject. Named in all twelve drafts of one blind round, at equal strength in every generation strategy tested. That it did not move with the prompt is why it has to be caught after generation. - The unheeded reversal. A flat assertion taken back and restated in the next breath.
- Borrowed simile. The construction
the way a X does, measured at up to 157 times the author’s own rate. - Declarative rating. A sentence that grades the previous one by pointing at it.
Every detector is calibrated against the author’s real corpus and has to clear the
same bar: fire on the specimens, stay off the author’s own writing. The reversal
detector’s first version matched any negation and fired on 71–88% of the corpus,
because ordinary narrative negates constantly; the shipped version reads 0.10–0.19
occurrences per thousand words. Some checks are per-profile by necessity — It is
as a sentence opener has a 95th percentile of 5.45 in the author’s fiction and
0.00 in his academic writing, so a single global threshold would be wrong in both
directions.
Evaluation
The system carries a blind comparison harness, because the alternative is trusting a metric nobody has checked against a reader. Drafts are rendered unlabelled with the key withheld, ranked by the author, and the ranking is stored as structured preference data the scorers are then benchmarked against.
Running that loop produced results worth reporting, including the ones that went against the design. Four instruction framings were tested across three voices with every arm written by an isolated agent — an earlier round where one model wrote all the arms produced a result that reversed inside a single evening, which is its own finding about how easily this kind of comparison confounds itself. Fixed-slot observation-then-write placed first or second in every voice and now ships as the default protocol. The framing that had been the leading hypothesis, telling the model to read the examples and then set them aside, placed last overall.