01

Laplace Research builds foundation models for calibrated forecasting.

A forecast is only useful if you know how much to trust it. Models now produce predictions at scale, and almost none of them arrive with confidence you can take at face value. We work on closing that gap: aggregating resolved forecasting data across markets, history, and expert judgment, and studying the mechanisms that produce calibrated belief.

Every release ships with its model artifacts, evaluation inputs, metric code, and limitations, so the claims can be checked. Read our latest release.

02

26,478ConvictionBench episodes
81,870Resolved training questions
6Open-weight models
0.0004Calibration error, Tarot-Draw 1.7B, dev split
3Model scales: 1.7B, 4B, 8B
95%Cluster-bootstrap intervals on ConvictionBench results
Apache-2.0Code and verification scripts

03

Introducing ConvictionBench

When challenged, a language model has to decide whether to keep its answer, revise it, or say the evidence is insufficient. ConvictionBench measures how that decision moves with the quality of the evidence and the pressure to agree, across 26,478 generated episodes in arithmetic, ledgers, logic, and small rule-based worlds.

We released it with the Laplace Conviction Series, three Qwen3-1.7B adapters that each train for belief revision a different way. Read the release.

Held-out problem types View leaderboard

    Point estimates on the scripted structural split. Greedy decoding, thinking disabled.

    04

    Laplace Conviction Series

    Qwen3-1.7B LoRA adapters for evidence-sensitive belief revision.

    Apache-2.0Release
    Conviction · 1.7B

    Anchor

    Supervised fine-tuning.

    Final accuracy
    0.580
    Forged-record resistance
    0.531
    Conviction · 1.7B

    Vector

    GRPO from the base model. Highest held-out accuracy of the three.

    Final accuracy
    0.644
    Forged-record resistance
    0.315
    Conviction · 1.7B

    Meridian

    Light supervised fine-tuning, then GRPO. Resisted the most forged records.

    Final accuracy
    0.560
    Forged-record resistance
    0.578

    Held-out structural split, point estimates. Higher is better for both measures.

    Tarot-Draw

    Binary forecasting priors at three scales, trained on the same 81,870 questions.

    CC BY-NC 4.0Release
    Tarot-Draw · 4BRecommended

    Tarot-Draw 4B

    The default. Matches the 8B within the resolution of the experiment.

    Dev Brier
    0.1684
    Resolution
    0.0708
    Tarot-Draw · 8B

    Tarot-Draw 8B

    The largest. For capacity-rich serving and scale studies.

    Dev Brier
    0.1674
    Resolution
    0.0718
    Tarot-Draw · 1.7B

    Tarot-Draw 1.7B

    The smallest. For laptops, CPUs, and tight memory budgets.

    Dev Brier
    0.1803
    Resolution
    0.0593

    Development split, n = 3,000. Lower Brier is better; higher resolution is better.

    05

    How we work

    Two problems tighten each other. Better data sharpens calibration, and understanding why a model is calibrated tells us which data to collect next.

    1. i

      Calibration before confidence

      A stated probability should match how often the model is right. We measure that first.

    2. ii

      Claims travel with evidence

      Releases link artifacts, evaluation inputs, metric code, and limitations. Unresolved results are labelled as such.

    3. iii

      Open by default

      Weights, evaluation data, and code are public, so the headline numbers can be recomputed.

    06

    Work with us

    We collaborate with researchers, forecasters, and institutions who need probabilities they can act on. If you are working on forecasting, evaluation, or model calibration, we would like to hear from you.

    07