Every claim shows its receipt

When Sumi answers a question, builds a quiz, or writes a flashcard, the claim carries a citation to the specific passage it came from: the document, the page or slide, and the excerpt itself. In study mode, clicking a citation jumps you to that exact spot in your source so you can read the surrounding context yourself.

Reading your material is the step everything else depends on, so it is never a black box. When you upload a photo or document, Sumi shows you the text it extracted before that text becomes study material, and you can correct it. A fixed word in extraction is a fixed fact in every quiz built on it.

Every citation wears a label describing how strong the evidence really is. Labels are assigned by the system and never upgraded for presentation. If we only matched a passage after the fact, it says so.

• Retrieved

  This passage was pulled from your material before the answer was written. The answer was grounded in it, not decorated with it afterward.

• Verified

  This claim was checked against the current version of your source after generation, and the source supports it.

• Related

  This passage was matched to the claim after the fact. It is relevant context, and it is labeled honestly as a match, not as proof.

• Unsupported

  We checked, and the current source does not support this claim. We show you that instead of hiding it.

• Unverified

  The evidence is stale. Either the source changed after this was generated, or the check has not run yet.

A quiz with a wrong answer key is worse than no quiz. Generated questions and their keys are support-checked against your source, and items that fail the check are flagged rather than quietly served.

When your material does not cover something, Sumi says so. Tools return “nothing found in your sources” instead of inventing filler. An honest empty result is recorded and stands, rather than being silently retried until something plausible appears.

AI models make mistakes; that is exactly why the receipts exist. Sumi does not present generated text as authoritative, does not grade you, and does not pretend confidence it has not earned. Its calibration readings describe what you can recall today, which is not the same thing as a grade. When the honest answer is “this is unverified,” that is what the label says.

How your work is stored and used is covered separately in the privacy policy.