de-AI Gate
A deterministic, read-only content-honesty gate. Two HARD scan modes:
- Served-HTML scan (
--html): char tells (em/en-dash, ellipsis, invisible unicode) + 12 high-precision AI phrase tells, after stripping CSS/comments. - Record-source scan (
--records): typographic-substitution corruption (curly quotes/apostrophe, no-break space, ellipsis char) on authored JS/JSON/TS spine - the check that catches AI-smuggled curly quotes a served-HTML gate misses.
Structural heuristics (sentence-length sigma, tricolon, transition scaffolding, watchlist clusters) report as advisory WARN and never change the exit code.
Invoke
# from the skill directory (or point node/python at its absolute path)
# served content
node cli.mjs --html ../my-site/public/articles ../my-site/public/guides
# authored records
node cli.mjs --records ../my-site/data/
# Git-Bash / Python surface (HARD tells only)
python gate.py --html ../my-site/public/ --records ../my-site/data/
Exit 1 on any HARD violation (use --warn-only for non-blocking baseline mode). Run --html and --records as separate invocations (or separate flags); mixing record and html targets in one call is not supported.
One catalog, two runtimes
All tells live in catalog.json. engine.js (JS) and gate.py (Python) both read it, so the tell definitions are single-source and never fork. Add a new tell once, in the catalog. Note: JS and Python regex word-boundary handling differs on text where a non-ASCII letter is directly adjacent to a tell word (no separator) - a near-zero case in real prose; both engines otherwise agree.
Scope
Gate reader-facing prose: README/docs, published pages, skill instructions,
release notes, About text, UI copy. Skip detector code and test fixtures (they
contain the glyphs they detect), commit messages (gate the shipped docs, not
history), and verbatim third-party quotes. Full scope rules in README.md.
Safety
Read-only file scan. No SessionStart/auto-run hook, no credentials or cookies, no network or egress, zero dependencies. See README.md.