ProofBridge

What it does

Every match deserves a second opinion. ProofBridge helps companies find research partners and understand the evidence behind each recommendation. It searches the official synthetic dataset of 1,047 researchers, then uses three Mozaik agents to propose a partnership, investigate counter-evidence, and independently review the result. Proposals include source-linked evidence, participation constraints, and a bounded pilot plan; unsupported recommendations can be revised or withheld. The public demo offers interactive replay of real saved runs, including three approved research proposals, with evidence inspection, review history, and version comparison. New analyses run locally with the user's own API key.

How the agents run concurrently

Each Run uses an isolated worker with a real Mozaik 4.2.0 Runtime. Proposer and Counter-evidence execute concurrently in separate Agent Contexts: one develops a partnership proposal while the other investigates limitations and contradictions. They share read-only access to the same challenge, fixed candidate pool, and source snapshot. A programmatic coordinator manages shared run state and joins their validated outputs. Critic then reviews the exact draft, counter-report, and evidence snapshot in an independent Context. It can approve, request one targeted revision, or return insufficient evidence. A revised draft receives a fresh review. The parent process controls request permits, cancellation, deadlines, and publication. An Agent's approval alone cannot publish a proposal. Persisted traces include native tool events and HTTP request intervals; the three approved runs show measured client-side request overlap of 4.943, 10.355, and 8.052 seconds. The public demo replays these recorded executions without issuing new model requests.