What it does
Evidence to Pilot helps coordinators turn a company's technical challenge into a research collaboration proposal for discussion. A TypeScript CLI uses four Mozaik agents: Company Analyst, Research Profiler, Proposal Writer and Critic. Local search links researchers to works by record IDs. Each brief contains source quotations, a proposed three-month experiment, baseline, decision criteria, requested resources and a meeting agenda. The recorded provider run in runs/live-8 accepted C001-CH1/R1008, C025-CH1/R0637 and C036-CH1/R0976 on fictional organizer data. The proposals cover industrial-network anomaly detection, circuit-board emission checks and raspberry packaging. Each required one Critic-requested revision. The run records 22 provider attempts and peak concurrent inference of 2; summary.json, trace.jsonl, briefs and evidence ledgers are preserved in that folder. Acceptance means mechanical and model checks passed, not partner consent or demonstrated feasibility. Runs in runs/live-1 through runs/live-7 preserve earlier failures and one accepted proposal; these development runs are not an accuracy benchmark. The pilots were not carried out, and EINSTEIN/Stars integration remains proposed. The system does not score people's worth. Judges can start with npm run demo after npm ci: small authored fixtures exercise the real runtime and correction path offline, using scripted model responses and no credentials.
How the agents run concurrently
One Mozaik runtime hosts four agents and an observer. For each challenge, Company Analyst and Research Profiler subscribe to challenge.opened and start separate runLoop calls without waiting for each other. They share the dataset, search index and retrieved work/researcher ID pairs. The observer validates their outputs before publishing findings to shared state and emitting findings.ready. Writer starts only when both findings exist; Critic then reviews the draft. Invalid outputs return to their producer, and Critic revision requests return to Writer through bounded correction loops. Challenges are processed one at a time; the parallel work is the two analysts within each challenge, not all four roles at once. All roles share a request-concurrency limiter and rate, attempt and estimated-cost budgets. In runs/live-8/trace.jsonl, the initial analyst loops overlap by 5,901 ms for C001-CH1, 6,729 ms for C025-CH1 and 6,640 ms for C036-CH1. runs/live-8/summary.json records peak concurrent inference of 2 and 22 provider attempts. Each proposal received one Critic-requested revision before acceptance; the trace also records four producer-validation send-backs. These are recorded overlaps, not measured speed-up against a sequential baseline.