01

Laplace Research

We build foundation models for calibrated forecasting and study why their confidence can be trusted.

02

Calibration before confidence

A forecast is useful only when you know how much to trust it. Models now produce predictions at scale, and almost none of them arrive with confidence that can be taken at face value. Laplace Research is an attempt to close that gap.

The goal is not simply better point estimates. It is to make model confidence measurable, reproducible, and useful enough that human judgment can be spent on choosing which questions matter.

03

Two connected problems

  1. i

    Forecasting data

    We aggregate resolved forecasting data across markets, history, and expert judgment. A model can only learn calibrated confidence from outcomes that are well specified and consistently resolved.

  2. ii

    Internal mechanisms

    We study the mechanisms that produce calibrated predictions. Understanding why a model is calibrated helps identify which data to collect next and where confidence should fail.

Our first public release was the Tarot-Draw family, a controlled study of model scale in binary forecasting, followed by ConvictionBench.

04

Founders

Bruce Nshuti Hirwa

Co-founder and Chief Technology Officer

05

Get in touch

For research, collaboration, and general inquiries, write to the address below.

Locations
San Francisco, CA and New York, NY
Code
GitHub