Analyze. Rewrite. Re-analyze.
Prove it or say so.

Every humanizer runs your text through a model. Offbeat is different in what happens before and after that call.

1. The analyzer reads first, on your device

Before anything is sent anywhere, a deterministic engine measures the draft: sentence-rhythm variance and its distribution, inflation vocabulary, stock phrases, punctuation habits, structural fingerprints. Twenty-plus checks per language, each with named evidence. Deterministic means the same text always gets the same score, so a number you saw yesterday still means something today.

The English rules descend from a scanner that is byte-for-byte verified against its reference implementation. The Arabic rules were derived from real corpora: three million words of human Arabic against eight million words of machine Arabic, with every rule surviving a false-positive audit before it shipped. Rules that failed the audit are recorded as killed, not quietly kept.

2. The brief targets what the analyzer found

The rewrite instructions are built from the findings: these exact phrases, this rhythm problem, those punctuation restores. Fix these, leave the rest. The model never gets a vague "make it sound human". Arabic gets its own brief, written in Arabic, that tells the model to chain clauses with simple wa- the way human Arabic actually does, and to leave the writer's dialect alone.

3. Several drafts compete, from more than one writer

The loop asks for multiple candidates at different temperatures, and the drafts come from more than one model, so a single model's bad day never decides your result. Deterministic guards disqualify any draft that dropped a number, damaged protected content, drifted in length, or mixed in characters from a foreign script, before scoring even starts. The analyzer from step one then ranks what survives, and only drafts that actually reduce the flagged patterns can win: a rewrite that changes your words but keeps the tells is refused, no matter how it scores. The committed evaluation runs behind every one of these choices live in the repo.

4. Protected content never enters the model

Code blocks, inline code, URLs, and front matter are replaced by placeholder tokens before the rewrite and verified back into place after it, byte for byte. A rewrite that loses a placeholder is rejected automatically.

5. The reviewer, the repair, and the honest answer

The winning draft still has to convince a reviewer: a separate model, never the one that wrote the draft, reads your original and the rewrite side by side and vetoes any lost fact, any invented detail, any broken grammar. A veto sends the loop back to work, first with a repair written from your original with the reviewer's findings spelled out, then with the smallest possible edit that removes only the flagged patterns. Real progress that leaves a few patterns in place ships with a notice that says exactly that. If the reviewer cannot be reached, the result says the rewrite went out unverified. And if nothing preserves your meaning, your text comes back unchanged and the result says so. A tool that can never say "your version was better" is advertising, not measurement.

The scorecard is the product

What you get back is the before-and-after table, check by check, plus a word-level diff of every change. You can verify the improvement, disagree with a specific check, or download the full report as JSON. That artifact is the whole argument for Offbeat: here are real ones.

Also speaks editor

The same engine ships as a CLI (offbeat scan exits nonzero on failing prose, so it works as a CI gate) and as an MCP server for agent hosts like Claude Code, where your editor's own model does the rewriting, so it costs nothing beyond the model you already run. Install steps for every surface live in the guides.