Delphi

Demo video

What it does

Three modes on one architecture. Forecast mode produces a probability with confidence interval. Pretest mode returns positive / neutral / negative sentiment for a tagline, message, or product concept. Stress-test mode surfaces the strongest objections to a decision before you ship it.

Every run returns: a headline number, ±1σ band, the strongest reasons for and against drawn from the agents' own reasoning, where demographics diverged most, and one striking outlier quote — all synthesized into a publishable paragraph by a final Gemini call.

Architecture

• N parallel Gemini 3.5 Flash sub-agents, each with a persona-specific system prompt sampled from US Census–aligned demographic axes (age, region across 9 Census divisions, education, income, occupation, belief axis).
• Native Google Search grounding on every agent — they read live news as they reason, not from a stale embedding index.
• Structured JSON output via Pydantic schema — the model picks position + confidence + reasoning, it cannot drift off format.
• Aggregator computes mean, ±1σ, distribution buckets, and a per-demographic-axis breakdown — deterministic Python, no LLM in the critical path.
• Final synthesis call — one Gemini call after aggregation produces a WSJ-style narrative, reasons for / against, a demographic split sentence, and a curated outlier quote.
• News-shock re-run — same personas, augmented context, re-reason in ~30s. The synthesis paragraph rewrites itself to cite the new information.
• Production cap of N = 20 for sub-90-second runs at conversational latency. Stress-tested end-to-end up to N = 200.

Validation

Not yet calibrated against real polling baselines. Synthetic forecasts have not been benchmarked against Polymarket, Pew, Gallup, or the Good Judgment Project. That's the next research piece — backtest harness against historical events scored by Brier loss against the GJP median.

Try it

The wedge

Forecasting is one use case. The same primitive pre-tests marketing campaigns at population scale, war-games policy decisions before they ship, runs synthetic juries for trial-message testing, and models behaviour for public-health interventions. Anywhere you need to ask how will the population react, this is the substrate.

Until Flash 3.5's cost × speed × intelligence frontier landed, hundreds of grounded reasoning agents in parallel at conversational latency was economically impossible. The cross-model comparison above is the evidence: the previous generation collapses on the same prompts.