Verification
Generated study material is cheap to produce and expensive to trust. This page states the guarantee precisely enough that you can hold it to account.
A card generated from material you supplied stores the exact span it came from, and that span is validated as a literal substring of the source before the card is accepted. A card generated from a topic prompt has no source, and it is labelled unsourced. There is no third case where a card implies a citation it does not have.
When a card is generated from your material, the cited span is checked as a literal substring of the ingested source chunk. Exact matching is strict on purpose: the failure it catches is a quote that was paraphrased or invented during generation, which is precisely the failure that a reader cannot detect on their own.
Some genuine quotes fail exact matching for uninteresting reasons: collapsed whitespace, a ligature, a line break inside a sentence. Those cases escalate to an entailment check rather than being waved through or discarded. The escalation path exists so that strictness does not quietly become a reason to loosen the rule.
Sources change. Stored spans are re-checked against the live source over time, and a quote that no longer matches is flagged as drifted on the card quality report. A drifted card is not silently deleted and it is not shown as trustworthy; it is shown as needing a look, which is the only honest state for it.
Cards generated from a topic prompt have no document behind them, so there is nothing to verify. Those cards are labelled unsourced. This is the part of the claim people usually skip: the guarantee is not that every card cites a source, it is that a card which cannot cite one tells you so on its face.
It confirms that a quote is really in your source. It does not confirm that your source is correct, current, or the edition your exam is set on. A card built from a superseded standard will verify perfectly and still be wrong for your purposes, which is exactly why the passage is shown rather than summarised.
It also does not evaluate whether the question is a good question. Card quality is reported separately, and weak items are flagged for review rather than being treated as verified simply because their quote checks out.
Spaced repetition works by rehearsing an answer until it is automatic. That is a powerful mechanism pointed at whatever you feed it, including a fabricated fact. A wrong card in a review schedule is worse than no card, because you will practise it until you are confident in it.
That is the reason grounding is checked at generation time rather than offered as a link for you to inspect later. By the time you notice, you have already rehearsed it.