cite-gate

A small checker for scripts written with AI help. It makes every factual item in a script carry a quote that appears word for word in a source file, and it flags prose that copies the source instead of retelling it.

It doesn't decide whether a claim is true. It decides whether you can point at the sentence your claim came from, and whether that sentence actually exists.

$ python -m cite_gate check examples/gettysburg/claims.json
5 items · 91 words · 2 failures · 1 warnings
  FAIL [number-not-in-quote] number(s) ['1776', '87'] not in the quotes
  FAIL [quote-not-in-source] quote not found in 'gettysburg': 'this nation shall endure forever and ever'
  warn [copied-not-retold] 26 consecutive words shared with the quotes; retell, don't copy

The example fails on purpose: its last three items are mistakes that a language model makes all the time (a number it worked out itself, a quote it half remembered, a paragraph lifted from the source).

Why

Language models write fluent prose with confident details. The details that go wrong are the small ones: a year that's off by one, a name added "for context", a quote paraphrased into something the source never said. Reading the whole script against the sources by hand doesn't scale. This tool turns that reading into a gate you can run on every change.

What it checks

Check	Result
Each quote exists in its source file (curly quotes, dashes and spacing normalised)	failure
Each quote is at least 20 characters	failure
Every number in the prose appears in that item's quotes (1,000 = 1000)	failure
Source ids are plain names (no paths)	failure
"hundred / thousand / million" in the prose but not in the quotes	warning
A capitalised word mid-sentence that isn't in the quotes (possible unsourced name)	warning
11 or more consecutive words shared with the quotes (copying, not retelling)	warning
Optional era rule: the prose names an era but the quotes don't tie the item to it; month names or "X era" phrases not in the quotes; special letters in the prose	warning / failure

Exit code is 1 when there are failures. --strict also fails on warnings.

Input format

A claims file (JSON) and a folder of plain-text sources, one file per source id.

{
 "config": {"allow": ["Lincoln"], "copy_run": 11},
 "sections": [
  {"title": "The occasion",
   "items": [
    {"id": "opening",
     "text": "He began by looking back four score and seven years ...",
     "claims": [{"src": "gettysburg",
                 "quote": "Four score and seven years ago our fathers brought forth on this continent, a new nation"}]}
   ]}
 ]
}

sources/gettysburg.txt holds the source text. Keep your own copy of each source (with its URL and revision noted at the top), so a later run checks against the same text.

Config

Key	Default	Meaning
allow	[]	names accepted without a quote (the subject of the script, for example)
copy_run	11	shared-word run that counts as copying
min_quote_chars	20	shortest accepted quote
fold_accents	false	compare Tokaido in prose with Tōkaidō in the source
era	null	{"name": "Victorian", "tie": "\\b18[3-9]\\d\\b", "months": true, "letters": ""}

Use

Command line (Python 3.10+, no dependencies):

pip install .            # from a checkout; adds the `cite-gate` command
cite-gate check path/to/claims.json [--sources DIR] [--json] [--strict]

# or without installing
python -m cite_gate check path/to/claims.json

GitHub Actions:

- uses: actions/checkout@v4
- uses: cite-gate/cite-gate@v0
  with:
    claims: script/claims.json
    strict: "false"

Use from your agent (MCP)

cite-gate ships a small Model Context Protocol server, so an agent can check its own draft before it shows it to you. It runs locally over stdio, has no dependencies, makes no network calls and collects no telemetry.

{
  "mcpServers": {
    "cite-gate": { "command": "python", "args": ["-m", "cite_gate.mcp_server"],
                   "cwd": "/path/to/cite-gate" }
  }
}

Tools:

Tool	Input	Output
check_citations	claims (the format above) and sources ({id: full text})	failures, warnings, passed
check_numbers	text and quotes	numbers in the text that no quote contains

There's also a hosted endpoint at https://mcp.mlpc.co.kr/mcp (MCP Streamable HTTP, same two tools, up to 3,000 characters of prose per call). It doesn't store your text; what it does log is listed in worker/README.md. For zero logging, use the local server above.

Would a hosted version or a review service be useful to you? Tell us in Discussions.

Limits

• The heuristics are tuned for English prose.
• A quote that exists is not a claim that is correct. Sources can be wrong, and an item can cite a real sentence that says something else. The gate makes the review smaller; it doesn't replace it.
• Proper-noun and quantity checks are warnings because they have false positives. Fix them or add the name to allow; don't learn to ignore them.

Example text

examples/gettysburg/sources/gettysburg.txt is the Gettysburg Address (1863), in the public domain, transcribed here from the Bliss copy. Small wording differences between Lincoln's manuscript copies exist; the checker compares against this file only.

License

AGPL-3.0-or-later. Copyright (C) 2026 MLPC Inc. (see LICENSE).