The Arbiter Financial (TAF) · Patent Pending
Before an AI agent's decision takes effect at your bank, TAF puts it in front of independent AI models from different providers. If they don't all agree, the decision stops and goes to a named person. Every step is sealed in a record your bank holds, and your auditors can check it without us.
Your AI model may be very good. But when it says "clear," nothing tells you whether it's right. A second opinion from the same model isn't a second opinion. When the provider ships an update, the decision boundary moves, and your audit trail says nothing changed.
Your bank chooses and seats multiple AI models from different providers. They can be self-hosted, in your own cloud tenancy, or reached through public APIs. Multiple copies of one model are not a panel.
A decision goes forward only when every model agrees. Anything else is parked for a person.
The parked item goes to a named supervisor, who releases it, blocks it, or escalates it to another named person, with a written reason. A release carries a confirmation token. An item that expires is never released.
Where the models diverge, the divergence itself is the output.
TAF is software that runs inside your bank's own environment. Divergence IP is the vendor. We supply the software and operate nothing. The panel input travels from TAF to the models your bank chose and back again. We never receive, relay or store any of it. Your bank holds every record and both signing keys. The only things that ever reach us are a five-field chain head and aggregate figures your bank has approved for release.
Each panel verdict, each park, each human decision and each escalation is a sealed, hash-chained, signed record. An auditor can verify these records offline, using your bank's public keys and the auditor tool, without contacting Divergence IP.
The TAF data custody overview is available on request. The full technical edition, with sample sealed records, is available under NDA.
Pilots, partnerships and licensing: robert@divergenceip.com