The site read like an LLM because it was

You know the sound. Even cadence, tidy tricolons, a fondness for the neat reversal, every paragraph landing on the same soft beat. It is not wrong, exactly. It is just nobody. The failure is not in any single sentence; it is that the prose has no fingerprints. You cannot fix that by editing harder, because the editor doing the editing runs on the same defaults as the writer. I needed a second opinion that was not another model with the same priors.

The first decision was the load-bearing one, and it was about framing. I was not building an AI detector. Those try to answer did a machine write this, which is both unanswerable and beside the point. I was building a voice-fit meter, which answers something smaller and more useful: how close is this passage to how I write. Fit, not authorship. That distinction is what keeps the tool honest. It never accuses anyone; it measures distance from a profile, and the profile is mine.

A meter is only as good as its reference. If I calibrated it on my recent writing, I would be calibrating on the exact contamination I was trying to catch. So the reference corpus is old: essays I wrote by hand between 2010 and 2012, years before a language model could draft a sentence. That corpus is unambiguously human and unambiguously mine, which is the whole point. From it the tool distills a calibration profile: sentence-length variance, the words I actually reach for, the punctuation I lean on, the shapes I repeat. The bar for on-voice comes from that, not from a vibe.

Measuring is half of it. The other half is a rewriter that takes an AI-shaped paragraph and moves it toward the profile. It drafts several candidates rather than one, scores each on the meter, keeps the closest, and then red-lines the change so I can see exactly what moved. It rewrites prose blocks only. It never touches a quote, a code sample, or a number, because those are facts, not voice, and a voice tool has no business editing facts.

Once the rewriter worked on a pasted paragraph, the obvious move was to point it at the archive. The batch mode scans existing posts, ranks them worst-fit-first, and proposes a rewrite per prose block. Nothing applies on its own; every proposal waits for me. Applying one snapshots a revision, so a later reseed cannot quietly clobber the human edit. A vague backlog of drifted prose became a queue I could actually work down.

Here is the honest boundary. The meter can be confidently wrong, and the way it fails is instructive. A post about embeddings is full of words like vector and cosine and dimension. Those are subject words, but to a naive lexicon they look like unusual vocabulary, which reads as a strong voice signal. So a piece that genuinely sounds like me scored 52, dragged down because its own topic was being counted against it. The fix was a topic-vocabulary discount: words that belong to the subject stop counting as voice markers, and the same passage came back at 74. I would not trust the raw number on any technical post without that correction, and I would not trust the tool at all as a verdict on someone else's writing. It is a mirror for one person, calibrated once, and it is honest only inside those limits.

Not authenticity, which is not a thing a score can grant. What it buys is a fast, specific answer to a question I used to answer slowly and vaguely: does this sound like me, and if not, where. The meter turns a nagging feeling into a coordinate, and the rewriter turns the coordinate into a diff. Whether the diff is an improvement is a judgment, and that judgment stays mine. It is the same division of labor I keep everywhere on this site: the machine proposes, a human signs.