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's not wrong. It's just nobody. And the problem isn't any single sentence. It's that the prose has no fingerprints. You can't fix that by editing harder, because the editor runs on the same defaults as the writer. So I went looking for a second opinion that wasn't just another model with the same priors.

Everything rested on the first decision, which was really about framing. AI detectors ask did a machine write this - a question that's both unanswerable and pointless. So I skipped it. What I built instead was a voice-fit meter, and it answers something smaller and more useful: how close is this passage to how I write? Fit, not authorship. That single distinction keeps the thing honest. It never accuses anyone of anything. It just measures how far a passage sits from a profile, and the profile happens to be mine.

A meter is only as good as its reference. Calibrate it on my recent writing and I'd be calibrating on the exact contamination I set out 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 for me. That stuff is unambiguously human and unambiguously mine. That's the whole point. From it the tool pulls a calibration profile - sentence-length variance, the words I actually reach for, the punctuation I lean on, the shapes I keep repeating. The bar for on-voice comes from that. Not from a vibe.

Measuring is only half the job. The other half is a rewriter that takes an AI-shaped paragraph and drags it toward the profile. It writes several candidates instead of one, scores each on the meter, keeps whichever lands closest, then red-lines the change so I can see exactly what moved. Prose blocks only. It won't touch a quote, a code sample, or a number. Those are facts, not voice, and a voice tool has no business editing facts.

The rewriter already worked on a pasted paragraph. So I pointed it at the archive. Batch mode scans what's there, sorts it worst-fit-first, and floats a rewrite for every prose block - but nothing lands on its own. Each proposal just waits on me. Accept one and it snapshots a revision, which means a later reseed can't quietly clobber whatever I edited by hand. That vague backlog of drifted prose? It's a queue now, and I can actually work it down.

Here's the honest boundary: the meter can be confidently wrong, and the way it fails tells you something. Take a post about embeddings. It's stuffed with words like vector, cosine, dimension. Those are subject words. But a naive lexicon sees unusual vocabulary, and unusual vocabulary reads as a strong voice signal. So a piece that genuinely sounds like me scored 52 - dragged down because its own topic got counted against it. I added a topic-vocabulary discount. Words that belong to the subject stop counting as voice markers, and that same passage came back at 74. I don't trust the raw number on any technical post without that fix. And I trust it even less as a verdict on someone else's writing. It's a mirror for one person, calibrated once, honest only inside those limits.

Authenticity? No. A score can't grant that. What it buys is a fast, specific answer to a question I used to answer slowly and badly: does this sound like me, and if not, where. The meter turns a nagging feeling into a coordinate. The rewriter turns the coordinate into a diff. Whether the diff is any good is a judgment call, and that call stays mine. Same division of labor I keep everywhere on this site: the machine proposes, a human signs.