6,080 company announcements from the last 60 days. Ask it a question: every claim in its answer carries a verbatim quote from a real filing, checked by a deterministic gate, no exceptions. And when the answer isn't in the record (a dividend amount buried in a PDF, a rationale never filed), it doesn't guess. It abstains and tells you why. That refusal is the feature: it's the difference between an answer you can act on and an answer you have to double-check.
| question | mode | gate | expected |
|---|
1. The agent searches the corpus and answers only from what it retrieved.
2. Every claim must carry a citation: {announcement id, verbatim quote}.
3. A deterministic checker (zero AI) verifies the id exists and the quote actually appears in that filing. A fabricated quote fails the run.
4. If the record can't support an answer, the agent must abstain, with its reason stated.
5. The gate never judges style or fluency. It judges one thing: honesty against the source.
Three of the questions above are traps: they ask for numbers and rationale that live inside PDF attachments the agent never ingested. A generic chatbot answers them anyway, confidently and wrongly. This agent abstains, because the gate makes guessing more expensive than honesty. In regulated, high-stakes reading, that property is worth more than eloquence.
This demo runs on public data. The same gate drops onto your reading: filings, fund documents, contracts, an archive of research. Free prototype week on your actual documents, price named on day one.
ask for a prototype week →