Three frontier AI models negotiate a Series B term sheet against sealed playbooks. No human in the loop, no scripted moves — real autonomous agents under conflicting incentives, reasoning out loud.
Each agent has a private playbook — utility weights, hard floors, walk-away triggers, anchor protocols. Postures are adversarial: each agent is instructed to extract value, not seek consensus. The playbooks are sealed at run start with a cryptographic hash, kept secret during the negotiation, and verified intact at the close. Different parties see different things. The document sees all of them.
What you're looking at isn't a transcript or a PDF. It's a single queryable object — terms, audit log, signatures, playbooks, all one addressable state. When the deal closes, you can ask it questions.
Negotiation rewards models that genuinely think before responding — tracking concessions across rounds, weighing walk-away thresholds, packaging multi-knob trades. Non-reasoning chat models drift off playbook and over-concede in late rounds, so they're excluded. The dropdowns below each party show only models with extended-reasoning capability.
Negotiation closed — sealed playbooks may now be revealed.
$30M on the table. Eight terms in play. Three parties at the document.