Summit Simulator
Summit Simulator lets you test a meeting before you hold it.
Core trick
Flash-driven guest agents, built from Deep Research profiles, play each session while a separate judge model scores the outcome.
Tags
Description
Summit Simulator lets you test a meeting before you hold it. Organizers of high-stakes gatherings, like an AI-safety summit with governments, labs, unions and civil society, only find out if the agenda worked after everyone has flown home. Usually it didn't: powerful voices dominate the big sessions, people only talk to their own side, and the final deal leaves half the room out.
We rehearse the day first. Deep Research builds 10 realistic guests with real positions and red lines. Gemini Flash plays them through every session of a one-day summit, and a skeptical Flash judge scores the result: did views move, did everyone get heard, did rival sides agree on anything? Gemini 4 Argon reads the transcript, explains in plain English what went wrong, and redesigns the agenda. Then it runs again. A 3D venue replays each session, and Argon checks screenshots of it to see whether last round's fixes actually worked.
In three rounds the score went from 32 to 74 out of 100, and the visual score from 2 to 8.2. When you approve a plan, the Antigravity agent writes a personal briefing packet for each guest plus a calendar and PDF bundle.
README excerpt
# Summit Simulator **Gemini 4 Argon designs your AI-safety summit and proves the design works before anyone books a flight.** Built for the Gemini 4 Argon SF Hackathon (Oct 25, 2026), **More Dimensions** track. **▶ 1-minute demo video:** https://www.youtube.com/watch?v=8k2zHl0WKLk ## The problem An organizer of a multistakeholder AI-safety summit (think Bletchley Park / Seoul / Paris-style meetings) has one day and ten delegations that disagree on everything: binding regulation, open weights, compute controls, evaluations and liability, benefit sharing. The default agenda is keynotes plus breakouts grouped by constituency. It produces speeches, not agreements, and the least powerful voices (Global South civil society, labor) leave feeling unheard. **Who it's for:** the summit's program lead, who has to sign off on the run of show. ## What it does 1. **Deep Research builds the guest list.** One `deep-research-preview-04-2026` background job per stakeholder *role*: EU AI Office, US natsec, UK AISI, a Chinese policy think tank, a frontier-lab CEO, an open-weights company, Big Tech policy, a safety researcher, Global South civil society and labor. Each job produces a cited brief. Flash turns each brief into a persona: stances on 5 issue axes (-5..+5), red lines, tradeoffs, engagement style and institutional power. Personas are **role-based and grounded in public positions**, not psychological profiles of real people. 2. **Flash simulates the summit.** Each