Check what your AI cites, down to the page
Your teams already ask AI to read the documents you trust: incident reviews, standards, the books your leaders quote. The answers sound right and arrive with citations. Checking them is the hard part. A citation that opens a whole book, or a passage near the claim rather than on it, leaves the reviewer doing the work the tool was meant to save. In one of my tests, Gemini Notebook wrote that developers lose 20–40% of their day to context switching, and cited a passage that gives no such number.
I built grounded-forge to make that check quick. It’s an open-source tool (MIT) for assistants that answer only from sources you chose. In its verbatim mode, every citation opens the author’s own sentence, highlighted on the printed page, and code, not the model, places every quotation. It runs inside Claude Code, on a Claude plan you may already pay for.
What I measured
Between 31 October and 6 November 2026, I put the same practitioner questions to grounded-forge, Google’s Gemini Notebook and Open Notebook, a popular open-source notebook tool, over the same 31 public sources. Two AI models from different companies, Claude and OpenAI’s, checked every citation against the books. Here is how often a citation said what the sentence beside it claimed:
- grounded-forge: 86–96%, depending on the mode and the grader.
- Open Notebook, on the same Claude model, told to cite only what a passage says: 93%.
- Gemini Notebook, given the same instruction: 64–67%.
- Gemini Notebook, as it comes: 35–37%.
All 94 quotations in grounded-forge’s verbatim reports matched the source word for word, because code places them. Told to quote only exact words, Gemini got 9 of 11 right. grounded-forge’s answers also carried more of the library’s supported ideas than either notebook’s.
Open Notebook came close on citations, which suggests much of Gemini’s gap is the model and the instructions. But each of Open Notebook’s citations names a whole source, with no page or highlighted passage, so the reviewer still has to search for the line. In my test, that meant a catalogue of 266 practices.

Four answers per tool are too few for a firm rate. The two graders differ by up to 9 points, one of them is the Claude model that wrote grounded-forge’s answers, and no human has graded the citations yet. The full results cover where Gemini led, how I improved both tools between rounds, a private library, and an ungraded side check of Google’s newest model.
A fix you can use today
For most uses, Gemini Notebook will be adequate, and one paragraph of instructions roughly doubled how often its citations held up. Set it as a custom chat style:
Attach a citation only to a sentence that says what the cited passage itself says. Do not add numbers, percentages, causes, examples, or applications to the user's situation that the passage does not contain, and do not make the passage's claim stronger or broader than it is. When you apply a source to the user's case, or draw your own conclusion, put that in a separate sentence that starts "My inference:" and carries no citation. Put words in quotation marks only when they appear word for word in a source. If no passage supports a point, say that no source in this notebook covers it rather than citing a nearby passage. Keep the answer in the chat.For work where a wrong citation is costly, grounded-forge is worth serious consideration, because every citation in a verbatim report opens the exact passage on the printed page.
How it works
A general-purpose model gives you the most-published answer, and the most-published writer is rarely the original thinker (a corollary to Larman’s Laws of Organizational Behavior). So you pick the sources you trust. A model reads each one in full and writes a reference with every claim traced to the text, and a fresh model audits those claims against the source before anything ships.

The audit isn’t a truth oracle. A fresh audit by a different model family found 97.45% of 9,364 reference claims free of hard errors before repair. It lets you check that the assistant sticks to the sources you chose. Whether those sources are right about the world stays your call.
In a verbatim report, the model writes the argument and picks the passage each point rests on, and code places the author’s sentence and records its page. (KR Labs’ open-source Verbatim RAG and Invetech’s “deterministic quoting” rest on the same idea.)
What you can do with it
Once your sources are in the library, you can use it three ways:
- Ask it questions while you work. In Claude Code, it answers from the sources you chose and cites them.
- Write verbatim reports. For a memo going to a decision-maker, verbatim mode builds the answer around the authors’ own sentences, each one a click from its place on the printed page. Here’s one, on Marine Corps doctrine and management writing.
- Build assistants for one job. One build command compiles the same audited library into an assistant per job. AAR Maker is one: a coach for after-action reviews that you can add to Claude as a connector (aar.groundedforge.com/mcp).
Throughout, I put the model where it earns its place and keep code where consistency, audit and cost matter.
Where it wins, and where it doesn’t
It costs more up front. Reading every source in full, under audit, takes about 1% of a week’s allowance on a Claude subscription for a 70,000-word book, and each verbatim report takes 13–23 minutes, unattended. Gemini Notebook needs nothing but the files uploaded. The trade pays off for a library your team will question often, on work where a wrong citation costs more than the reading.
On well-known public books with descriptive file names, a plain model reading the files did as well as the prepared library in my tests. Grounding earns its keep on your own material, the engagement documents and incident histories no model was trained on. Placing quotations by code also didn’t make citations hold up more often than grounded-forge’s standard mode, and in one round 20 placed passages lost words before a fix. Its gain is that each citation is quick to check.
For builders
The demo library ships 31 open sources prepared for five jobs, compiled into five assistants from one build command. Fork it, strip the demo content in one command, and add your own sources. The code is MIT, and everything is plain markdown, so your library leaves with you. The repo is at github.com/chrisgagne/grounded-forge.
If you want this for your own documents
If your team answers from documents you’d stake a decision on, I can build a library from them and test its citations against your own questions, the way I tested it here. It runs in your own Claude Code, so your documents stay in your repository and go only to the model provider you already use. Tell me which documents and which questions, and I’ll scope it with you.