Personal
The de-LLM loop — hunting the machine tells
Part of the technology exposé. The closed editorial loop that drives prose toward "no obvious craft issues or LLM tells" — and the duplicate-tell eliminators at its core.
A model's prose has a fingerprint: the spaced em-dash, "almost smiled," the thesis stated on a loop, the flat even rhythm where every sentence weighs the same. Readers feel it as "this was written by a machine" without being able to name it. The de-LLM loop's only job is to find those tells and kill them — and to learn each one permanently, so the prose quality ratchets up instead of drifting.
flowchart LR
R[Cold-read + craft audit<br/>find craft issues / LLM tells] --> X[Re-incorporate into engine<br/>prompts · style guide · tic scanners]
X --> F[Surgical edit pass<br/>human in the loop]
F --> S{Tic scanners<br/>+ scorers}
S -->|tells remain| R
S -->|only creative changes left| DONE([ship])
Three components chasing each other
- A brutal cold-read agent (sentence layer) and a structural craft audit (the layer above —
voice homogenisation, gravitas inflation, over-polished action, reveal/reaction order) find the problems model prose falls into.
- Each finding is re-incorporated into the engine — the prompts, the style guide, and a set of
deterministic tic scanners that count the specific machine-tells against falling targets.
- A surgical, human-in-the-loop prose pass fixes them. Then the loop runs again.
The duplicate-tell eliminators (deterministic, free)
The "duplicate eliminator" is really a family of scanners, each killing a different layer of repeated machine pattern. They are pure pattern-counters — no model, no cost — and they run on every build:
| Scanner | The tell it eliminates |
|---|---|
prose_tics | the sentence-layer tells — the spaced em-dash, "almost smiled," the stock gesture-beats a model reaches for again and again |
prose_thesis | the semantic tell — the book's theme stated and re-stated on a loop, as if the reader might miss it |
prose_evenness | the deepest and subtlest tell — even register, where every sentence carries the same weight and rhythm, so nothing lands. The cold-read names this the hardest to grep; it's made measurable here. |
prose_cadence | rhythm and sentence-length variance — the music under the prose |
voice_audit | character-voice homogenisation — when everyone in the cast starts to sound the same |
Each one turns a vague "this feels like AI" into a number against a target, so a revision can be judged on whether it actually removed the tell or just moved the words around.
Why the loop matters
The system learns from its own failures. A tell found once becomes a guardrail that catches it forever after. That's the whole trick: the prose doesn't depend on the model getting better — it depends on the scanners getting more complete, which they do every time the cold-read finds something new.
← Back to the technology exposé · graded against targets by NovelBench.
Craft Library · Workshop · View this document on GitHub · Write with us