AI writes with confidence.
Shisai checks it against the record.
A verification layer for AI-drafted claims. It checks whether what the AI wrote is backed by the document it cites, shows you the exact passage, and produces a record a compliance officer can file.
No model makes the judgement. The same claim against the same record gives the same answer every time — and the answer comes with its evidence.
On your side
Most AI-governance tools record what the AI said so the firm can hold someone accountable later. Shisai checks the claim before it matters — for the person who can still fix it. The compliance file still exists; it's the only version where you saw it first.
They watch you. We check the work.
Honest by construction
Every verdict names what was found: faithful, not supported, contradicted, source nonexistent, unreadable, calculation wrong. A claim it couldn't check is reported as unchecked — never padded over, never guessed.
It never tells you something is fine when the record doesn't say so. That rule halts everything else when it's broken.
Who it's for
Firms professionally punished for being wrong. First: compliance-led investment firms that must show a regulator the substantiation behind AI-drafted material. Second: products that draft from documents and need "traceable" to be literally true.
Works inside the AI tools people already use (a connector for Claude Desktop and Claude Code today), as a sweep across an organisation's AI output, or as an API.
Early, and looking for the first firms
Shisai is in testing — AI agents use it unprompted, and one has already withdrawn a claim when shown the evidence. We're opening a small number of design-partner pilots with regulated firms and document-drafting products.