Delphi
Delphi · synthetic populations as a computational primitive.
Core trick
Runs parallel grounded persona agents built from census demographic axes and aggregates their answers into a probability with confidence interval.
Description
Delphi · synthetic populations as a computational primitive.
You ask any question — Will the Fed cut rates in Q3? Pretest this tagline. Stress-test this decision — and a swarm of Gemini 3.5 Flash sub-agents, each role-playing a different American persona generated from real US Census demographic axes and grounded in live web, reasons in parallel.
In ~60 seconds you get back: a probabilistic forecast with confidence interval, the strongest reasons for and against drawn from agents' own reasoning, the demographic axes where groups diverged most, and a striking outlier quote — all synthesized into a Wall Street Journal–style summary by a final Gemini call.
Live-validated on stage at N=500 with a +66.2 percentage point shock response (13.3% → 79.5% when CPI surprised). Validation across 27 automated tests, an adversarial persona-stability harness (4.92/5 in-character, 0% drift to centrist mean), and a cross-model comparison that showed Gemini 3.5 Flash holds 100% per-agent success where Gemini 2.5 Flash collapses to 12.5% on identical prompts.
Until Gemini 3.5 Flash made hundreds of parallel grounded reasoning agents economical at conversational latency, this category did not exist. Forecasting is the wedge — the same primitive powers marketing pretests, policy war-gaming, synthetic juries, and behavioral pre-mortems.
README excerpt
# Delphi **Synthetic populations as a computational substrate.** Built on Gemini 3.5 Flash for the Google I/O Hackathon, 2026-05-23. You ask any question — *Will the Fed cut rates in Q3? Pretest this tagline. Stress-test this decision* — and a swarm of Gemini 3.5 Flash sub-agents, each role-playing a different American persona generated from US Census–aligned demographic axes and each grounded in live web, reasons in parallel. In ~60 seconds you get back: a probabilistic forecast with confidence interval, the strongest reasons for and against drawn from agents' own reasoning, the demographic axes where groups diverged most, a striking outlier quote, and a Wall Street Journal–style synthesis paragraph — all produced by one final Gemini call after aggregation. It is not a chatbot. It is not a copilot. It is a new primitive. --- ## Headline result Live, on stage, **N = 500** synthetic Americans, the question *"Will the Fed cut interest rates in Q3 2026?"*: | | Before shock | After shock | Δ | | ------------------------ | -----------: | ----------: | ------------: | | Headline | 13.3% | **79.5%** | **+66.2 pp** | | ±1σ band | [7%, 19%] | [73%, 86%] | tighter | | Bullish bucket (≥ 0.8) | 0% | 62% | full flip | | Per-agent success | 363 / 500 | 409 / 500 | improved | Shock injected: *"Surprise May CPI prints 2.1%, well below forecasts."* The pos