search the future

OpenSuperforecaster

Turn a hard question into cited probabilities you can audit.

A self-hosted forecasting appliance. It fans your question out to parallel research agents, aggregates their findings into an inspectable probability estimate, and keeps every source, attempt, trace, and benchmark score on your own machine.

06 Forecast modes
03× Parallel agents
04 Local start steps
100% Local trust chain
01

Trust Chain

Don't trust a score — inspect the chain that produced it. Every run keeps its work.

01

Ask

Pose a binary, numeric, date, categorical, thresholded, or conditional question with an explicit resolution rule.

02

Fan out

Parallel Smithers agents gather evidence independently, so disagreement stays visible instead of averaged away.

03

Synthesize

Findings aggregate into a probability or distribution with rationale, caveats, and inline source citations.

04

Audit

Sources, attempts, trace bundles, and benchmark scores persist as inspectable local records you can replay.

02

Architecture

One local pipeline: question in, cited probability out, with every intermediate step persisted.

Open Superforecaster architecture diagram showing question input, agent fan-out, source research, synthesis, probability, artifacts, traces, benchmarks, and local store
01 · Input

A question fans out to parallel agents that research sources, then synthesis emits a probability and forecast artifacts.

02 · Research

Agents pull from the open web, arXiv, GitHub, reference data, and more — every citation is captured in the source bank.

03 · Trust chain

Traces, benchmark cases, and a local store back the answer so any run can be replayed and re-scored.

03

Cockpit

The real product surface: compose a question, launch a source-backed run, inspect the record.

local appliance

Forecasts are records, not chat replies.

Ask a question and watch researchers gather evidence in real time. Every run becomes a local object: the question, its sources, individual attempts, the aggregate answer, artifacts, and the full workflow trace.

localhost:3000
Open Superforecaster composer: a forecast question ready to launch
localhost:3000/runs/osf-resolution-smoke
Run detail: a binary forecast output with outcome bar, artifact rows, sources, and a scored forecast ledger
run detail

Every answer opens up.

Each run detail page surfaces the emitted probability, the artifact rows behind it, the source bank, and a forecast ledger of attempts, aggregates, and Brier + log scores. Export the trace bundle, CSV, or Parquet for deeper review.

CSVParquetTrace bundlesArtifactsDuckDB tables
04

Forecast Modes

Six typed forecast modes, each with a dedicated renderer instead of a wall of freeform text.

01

Binary

P(event) as an explicit yes/no probability.

02

Numeric

A point estimate with a percentile range over a continuous quantity.

03

Date

When an event resolves, expressed as a distribution over time.

04

Categorical

Probabilities across a set of mutually exclusive outcomes.

05

Thresholded

The odds a tracked metric crosses a specific threshold.

06

Conditional

How one event shifts the probability of another.

05

Evaluation

Forecasting systems fail when their assumptions, evidence, or scoring setup stay hidden. This one makes them boringly visible.

benchmark lab

Score the workflow before you trust it.

Run benchmark and pastcasting loops against fixed evidence from the same operational lab that tracks diagnostics, recent runs, pending resolutions, and workflow health before a forecast workflow gets trusted.

localhost:3000/lab
Benchmark Lab dashboard with workflow launcher, diagnostics, recent runs, and benchmark status
01

Question

What exactly was forecast, and how does it resolve?

02

Sources

Which evidence moved the probability, and where did it come from?

03

Disagreement

Where did individual researchers diverge before synthesis?

04

Trace

Can the run be replayed from prompt to artifact bundle?

05

Score

Did the workflow survive local benchmark and pastcasting checks?

06

Action

What should be re-run before anyone trusts this in a real decision?

06

Run Locally

Open forecasting infrastructure should be inspectable, forkable, and runnable by default.

clone and start

Bring up the appliance on your machine.

Postgres, object storage, the web cockpit, and the worker come up from the repo with Docker Compose, so the full trust chain stays local and inspectable.

Agent subscriptions are mounted through one local auth root. Keep Codex as the default, or configure Claude Code, Pi, Kimi, and other Smithers CLI providers from the setup page after their profiles are logged in.

Next.js Local cockpit UI Postgres Product ledger MinIO Artifact + trace storage DuckDB Local analytics Smithers Durable workflows Smithers agents Configurable fan-out
local start
git clone https://github.com/ralfboltshauser/open-superforecaster.git
cd open-superforecaster
cp .env.example .env
mkdir -p data/agent-auth/codex/default
docker compose up --build