Live site: https://chndr-prksh.github.io/murmur/
Ask what happens next. Murmur researches a question with current news, forecasts the outcomes, and simulates how one million people react.
- Research and framing (
worker/): a Cloudflare Worker calls Claude with web search. Claude finds the closest past events, defines 5 to 8 checkable outcomes and 5 to 8 audience groups, and estimates a probability for each outcome. - Panel: the same brief goes to several independent Claude samples. Their probabilities are averaged, and the lowest and highest are kept as the range.
- Evidence tone (optional): Laya, a small open decision model, scores each source as critical, neutral or supportive, independently of Claude. Its reading is shown on the page and passed to the panel as one signal.
- Population simulation (
site/sim-worker.js): the per-audience rates (how many hear about it, how they split, how many pass it on) drive a spread model of 1,000,000 simulated people, run 20 times in the visitor's browser. - Record: every forecast is saved with its timestamp so it can be scored later.
The probabilities are model judgement and the population figures are a simulation. Neither is a measurement, and forecast accuracy has not been backtested yet.
| Path | What it is |
|---|---|
site/ |
Static front end, no build step. Deployed to GitHub Pages by .github/workflows/pages.yml. |
worker/ |
Cloudflare Worker that holds the API key and produces forecasts. |
Without a backend URL in site/config.js, the site runs on a hand-written sample forecast and says so on the page.
python3 -m http.server 4173 --directory sitecd worker
npm install
npx wrangler kv namespace create FORECASTS # paste the id into wrangler.toml
npx wrangler secret put ANTHROPIC_API_KEY
npx wrangler deployThen set ALLOWED_ORIGINS in worker/wrangler.toml to your site's URL, redeploy, and put the worker URL in site/config.js.
Laya needs its own host with Python; it cannot run inside the Worker. laya-server/ holds a container image for Google Cloud Run.
pip install "laya[serve]"
laya-serve # listens on port 8000Put that server's public URL in LAYA_URL in worker/wrangler.toml. If it is unset or unreachable the step is skipped and the forecast still completes.
Each new question makes one web-search call and several panel calls to Claude Opus 5.5. DAILY_LIMIT caps new questions per visitor per day and repeated questions are served from cache for six hours..
MIT