Veritaserum

Veritaserum is a shared team agent with private conversations when needed. People collaborate on one task and choose whether each message and reply is visible to the team or only to themselves. The server constructs the model’s context for that audience: private turns can use shared knowledge and the sender’s own evidence, while team turns receive shared conversation and explicitly approved contributions. When private evidence can help the team, Veritaserum proposes an exact statement for its owner to approve. That statement enters shared context without copying the underlying private conversation. The first integration is checkout incident recovery: support, engineering, and operations combine their perspectives, compare candidate repairs, and approve a sandbox action that must pass verification before it is applied. The project demonstrates successful collaboration across information boundaries, supported by inspectable planning inputs, versioned approvals, stale-plan rejection, and 55 automated tests. It is a working local prototype with synthetic data, not a production security certification.