Skip to content
Work with us

Research

Our research goes into Lemma, our reasoning engine, today, and into our own models next.

View publicationsRead the roadmap

Agenda

  • Verifiable reasoning

    Models that break problems into claims and prove each one, so conclusions can be checked.

  • Honest uncertainty

    Step-level confidence that tells people where a model is likely to be wrong.

  • AI for science

    Research agents that design experiments to falsify hypotheses, not confirm them.

  • Measurement

    Evaluations that grade the reasoning, not just the final answer.

Roadmap

Where each stage stands today.

Full roadmap
  1. In production

    Now

    The Lemma engine

    Our verification layer on leading foundation models, running in our product and in enterprise agents.

    • –Step-level confidence and review routing
    • –Step Validity evaluation on real work
    • –Full reasoning traces for every answer
  2. In development

    Next

    Our first model

    A model trained on verified reasoning traces, built to check steps more reliably and more cheaply than a general model.

    • –Trained only on data we have the right to use
    • –Measured on Step Validity before it ships
    • –Published with its evaluation results
  3. Planned

    Then

    Our models under the engine

    Theoratix models do most of Lemma's work, with foundation models wherever they are still the better choice.

    • –Same product, lower cost per answer
    • –Private deployment inside client networks
    • –Research published as we go
  4. Goal

    Goal

    The frontier of reasoning you can check

    The most reliable reasoning systems for decisions that have to be explained, measured in public.

    • –Leading results on public reasoning evaluations
    • –A safety framework applied to every release

Releases

Oct 2026

Lemma engine

Our reasoning engine, in production in our product and first agents.

Evaluations

Our own evaluations, and the public benchmarks we report.

  • Step Validity

    Theoratix

    Scores each reasoning step for correctness and for whether later steps depend on it.

  • Faithfulness probes

    Theoratix

    Tests whether the reasoning a model shows is the reasoning that produced its answer.

  • Calibration under shift

    Theoratix

    Measures whether stated confidence still matches accuracy on unfamiliar problems.

  • Public benchmarks

    External

    MATH-500, GPQA Diamond and SWE-bench Verified, run with published settings.

Posts

All news

Safety · October 9, 2026

Publishing our Reasoning Safety Framework

Publications

View all