Use cases
What the engine is proven to do today. Nothing on this page is aspirational: every capability below passes verification against controlled fixtures where every true answer is known in advance. No numbers yet, on purpose; those publish on the methodology page when the full proof run completes. Until then, capabilities only, and only proven ones.
Grounded answers with citations
Ask a question over your own documents and get the answer with its receipts: the exact lines it came from, in the exact files. Not a similarity score, a pointer you can check.
Honest abstention
When your documents don't contain the answer, it says "Not mentioned" and names what it looked for, instead of inventing something plausible. The silence carries as much information as an answer would.
Memory that updates over time
Facts change, and the memory keeps up without losing its past. The latest value wins the "now" question. The old value still answers questions about when it held. Nothing gets silently overwritten.
Connections across documents
Answers that no single document contains. One file says who leads the team, another says what the team owns, and the question that needs both gets answered by walking the chain, with each hop witnessed in its own source.
Counted and listed answers, with their set
"How many" comes back as a number plus the members it was counted from, each member carrying its own source. If the set can't be shown, the count doesn't get asserted.
More cards land here as the benchmark run proves more. When the numbers behind these capabilities publish, they publish under the rules on the methodology page.