Rewrite AI tells out of the text live, on the device

[peopleworks]

Your text on the left, the de-AI-ified version on the right, both rebuilt on every keystroke, with the score falling as you go. It runs off the rule-pack — no model, no network, no API key — which is what makes a live view possible at all. Every change is listed with alternatives to pick from and a one-click leave this one alone. Three strengths, from only the strongest tells to delete the empty intensifiers too.

The existing BYOK Humanize stays for what genuinely needs a model. Most of what it was doing, though, is a word swap the packs already knew.

The hard part was refusing to rewrite

A word swap breaks prose in ways that are obvious once you see them, and a tool that advises people on writing can't be the thing producing them. Each of these came out of reading the real output, not out of theory:

It would have produced	Now
delve into → "examine into the data"	declined — and alternatives don't save it: "look into" would give "look into into"
The report showcased → "The report show results"	offered, never auto-applied: one rule covers delve/delves/delving
it's not just a tool → "it's not a tool" — inverted	declined: the word holds up a flagged construction
a plethora of options → "a many of options"	declined: the "a ___ of" frame needs the determiner rebuilt
el panorama → "el situación"	declined or gender-matched — read off the article, not the ending, because panorama and problema are masculine despite the -a
a crucial step → "a important step"	corrected to "an important step"

Rhythm and rhetorical structure need real rewriting, so they stay in the recommendations — and the panel reports how many tells that covers rather than implying they were handled.

Replacements are data, not parsed prose

suggestion is written for a person ("mix, blend, range — or just name the thing"), so lexical rules now carry explicit replacements / delete fields. Inferring deletion from prose would have meant matching "usually deletable" and "suele sobrar" — and failing silently for any language the packs don't ship. Community catalogs written before these fields still work, but the fallback trusts nothing except an unmistakable comma-separated list: a lone term could be the replacement "use" or the description "muletilla", and guessing wrong rewrites someone's sentence into nonsense. A test keeps every built-in rule explicit.

Checks

• dotnet build clean; 125 tests pass (118 before this branch's additions → 121 → 125 as each guard landed).
• Verified in the browser in both languages. English sample 94 → 90, Spanish 91 → 85, and every substitution in both is grammatical and meaning-preserving.
• Two of the guards above exist because the first working version got them wrong on screen.

🤖 Generated with Claude Code