Skip to content
Work with us

The reasoning engine

Lemma turns a question into claims, checks each one, scores how sure it is, and sends what it is unsure of to a person. It runs inside our product and the agents we build for enterprises.

Watch it reasonRead the roadmap

What it runs on

Today Lemma runs on leading foundation models, chosen for each task, with our own verification and evaluation layer on top. Our own models are in development.

  1. Foundation modelsLeading models from several providers, chosen per task and swappable without changing the product.Partners
  2. Verification layerDecomposition, checks, confidence calibration and review routing.Theoratix
  3. EvaluationStep Validity, run continuously on real work, so reliability is measured rather than claimed.Theoratix
  4. Theoratix modelsModels trained on verified reasoning to do the checking more reliably and at lower cost.In development

How it works

  1. Decompose

    Breaks a question into claims small enough to check one at a time.

  2. Check

    Verifies each claim with code, a proof assistant, a search or a second model.

  3. Score

    Gives every step a calibrated confidence, so people know where to look first.

  4. Review

    Sends the steps it is least sure of to a person before anything counts.

  5. Record

    Keeps the whole trace: every claim, check and decision, ready to audit.

Capabilities

PythonLean 4SearchSQLYour systems

Tool use

Calls a code interpreter, a proof assistant or your own systems when reasoning alone is not enough.

  • Restate the claim0.99
  • Split into cases0.96
  • Pattern from examples0.62
  • Proof by induction0.99

Calibrated confidence

Flags the steps it is least sure of, instead of sounding equally certain about everything.

Verifiable reasoning

Every answer comes with a chain of steps that a reviewer can follow and check.

People in the loop

Low-confidence steps and high-stakes actions wait for a person to approve them.

Inside the trace

Lemma labels what each step is doing. Guesses are marked as guesses, and checked before they count.

Lemma · trace viewertrace_4f2a
  1. [claim]n² + n is even for every integer n.
  2. [rewrite]n² + n = n(n + 1).
  3. [unsure]Even for n = 1…5, so probably always even.
  4. [tool:lean]Prove: one of any two consecutive integers is even.
  5. [verified]So n(n + 1) has an even factor and is even.

Samples

Pick an example and watch the reasoning stream in.

Step-by-step reasoning that marks where it is unsure.

Question

What is the sum of the first 100 odd numbers?

    Illustrative examples of reasoning traces, not live model output.

    How Lemma reasons

    01/ 04

    Decompose

    Lemma breaks the question into smaller claims it can check one at a time.

    Is n² + n even for every integer n?n² + n = n(n + 1)confidence 0.99n, n + 1 are consecutiveconfidence 0.99One of them must be evenconfidence 0.98So n(n + 1) is evenconfidence 0.97
    Answer: yes, always even

    Illustrative example of a reasoning trace.

    Where it runs

    Lemma is not sold as a public API. It runs inside our product and the agents we deploy for organizations.