Renamed to "Humanize AI Slop Lyrics" - clearer about what it does (detect and humanize), and your install updates in place. The detector was retrained on a larger song corpus for better, fully repeatable accuracy. Same 100% on-device, no-account, no-network design; the humanizer works the same.

v1.0.0 — first public release.

What it does: scores how "AI" your Suno lyrics read (0-100%) and helps you humanize them. The Humanizer has three levels — Line (rebuild your single worst line), Rewrite (swap every cliche word), and Chaos (a full rework) — and it keeps your rhymes and hooks.

New in 1.0.0: a compact tap-to-cycle feedback panel, and the Humanizer now never makes a line read MORE AI than it started.

Runs 100% on your device — no account, no sign-in, nothing uploaded. It reads only the lyrics box on suno.com.

Smarter, gentler Humanize. The buttons now only rebuild lines that actually carry AI tells (known cliches or a high per-line AI reading) - lines that read human are never touched, and a repeated hook/chorus is recognized as song structure and left alone. If nothing can honestly be improved, the app now says so (and points at the structure) instead of editing something innocent. Each press generates up to 10 candidate replacement lines and judges them with the trained model plus six craft lenses, keeping the best one - and a press can still never make your score worse. Also fixes rare junk in generated lines (publisher-footer fragments and non-words from the training corpus are now filtered out).

The Humanize buttons are all-new. They now rebuild your most-AI lines as brand-new lines, written the way a freestyler writes: rhyme locked first, then the line built backward to match your song's theme and syllable count. "Humanize Line" redoes the single worst line per click, worst-first. "Humanize Rewrite" redoes the worst half of the song in one press and keeps the better half yours. Undo reverts either. The generator cannot use known AI cliches (they are excluded from its vocabulary), never copies real lyrics (4-consecutive-word guard), and every change is gated so the AI% can only drop. Still 100% on-device: no network, no LLM, no new permissions.

v0.5.1 — Bug fixes. The score now clears and recalculates when you move between Suno songs, refresh, or leave a song page (it no longer holds the previous song's number). The popup's paste-box tester now uses the same model as the on-page pill, so the two scores always match. Also runs on the Create page.

v0.5.0 — Retrained, smarter model. It's now weighted to how people actually use AI for lyrics, with a new "typicality" signal that measures how much a song is built from generic AI phrasing. New: it guesses WHICH AI likely wrote a song — Suno, ChatGPT, Claude, Grok or Gemini — shown only when it's confident. Fewer false alarms on human-written songs. Still read-only, on-device, and for fun.

v0.3 adds six on-device "craft lenses" (rapper, poet, psychologist, philosopher, storyteller, wit) — about 55 new text-only signals for richer, example-based feedback. Retrained on real consumer-AI sources so it now catches Suno-style lyrics it used to miss, while keeping human classics low. Also new: on-device non-English detection (an honest "English-only for now" note instead of a meaningless score), eye-rhyme detection, and a smarter feedback panel. Still 100% on-device, text-only, no network.

v0.2 swaps the hand-tuned cliche meter for a real trained model and adds a craft coach.

• The score is now a logistic-regression model trained on 3,805 AI songs vs 3,848 human-song metric vectors (bag-of-words + 79 text-only craft/stylometric features, ~85% cross-validated). The number you see is pure, temperature-calibrated P(AI) - no more blended guesswork.
• New craft-coach panel: 5 things the song does well, 1 creative move to try, 5 things to work on - every note quotes a real word or line from your song with a one-line fix.
• Smarter input handling: strips [Verse]/[Chorus] tags, leaked JSON and model scaffolding before scoring; instrumentals are detected and skipped.
• Still 100% on-device, text-only, no network - the model ships inside the extension. No embeddings, no LLM calls, nothing uploaded.