When you have an AI gather your notes and documents into a page, the sentences now often come with a source attached. A source at the end of a sentence makes it feel checked. But open the source and read the two side by side, and sometimes a number is different, or the source does not go as far as the sentence does. You can ask the same AI again, "does this sentence match its source?" What comes back is "yes, it does", and you cannot see which passage of the source it read to say so. Opening and comparing sentence by sentence stops being practical as the pages grow.
Graphium has a "source check" that compares the sentences of an AI-written page with the sources they cite. Open a claim (a key point taken out of a note or a document) or a topic (a page per subject), press Check sources in the banner, and the AI re-reads the source and gives a verdict. There are five verdicts: Found in source, Differs from source, Not in source, Couldn't tell and Source unavailable.
What I cared about is that the verdict itself can be checked. The verdict also comes from an AI, so I thought a bare "found" would only add one more thing to verify. So each verdict comes with a short reason, and with the quote from the source that the AI based it on when it returns one. When the source is a PDF or a Word file and the quote can be pinned to one place, that place appears under the quote, such as "Page 12" or "Paragraph 12". That place is not written by the AI. It is worked out mechanically from the original text that was read for the check.
I kept what it checks narrow on purpose. It only looks at whether the cited source says so, and does not judge whether the statement is true. The source is read again every time, so for a web page the comparison is with what the page says now. A topic is checked sentence by sentence, and if even one sentence comes out as "Differs from source", that is what the page's badge shows.
I wanted it to be enough to look only at the items with a worrying verdict, without comparing every sentence. Items judged "Differs from source" or "Not in source" collect in a Needs review list. The AI's verdict can be wrong, so a verdict alone does not hide an item from the lists. You open the source and read it, mark the item with Confirm if it holds, and fix only what you want to fix.
What a source says and what the world already knows are two different questions. The second one is handled by a separate feature, "world grounding", which compares a claim or an insight with Wikipedia and OpenAlex (an open database of scholarly papers). It places the claim as consistent with established knowledge, or as having counter-evidence, and it does not declare anything true or false either. I use it as a hint for where to dig deeper.
The source check needs an AI. You register at least one model in the desktop app first. Its details section is still marked Preview. Insights (generalizations across several claims) are not covered by the source check, because there is no original text to compare them with. When the source is a PDF, clicking the source name opens the page that holds the quoted sentence, but for notes and Word files it only opens the source and does not scroll to the sentence. World grounding is a preview feature that you turn on in Settings before using it.
How to read the verdicts and how to check many items at once are covered in the "World grounding" page of the manual. World grounding came in v0.8.0 (2026-05-21), and the source check in v0.77.0 (2026-09-17).
https://kumagallium.github.io/Graphium/manual/ai-grounding
Graphium is open source on GitHub.


Top comments (0)