Think-Reason-Learn
NewFounderBrain, our first pre-trained model

The decision harness for predictions

Language-model judgment at machine-learning speed. More accurate, explainable and almost free to run. By Vela Research and the University of Oxford. Accepted at IEEE, NeurIPS'26 Workshop and more.

Will this startup become a unicorn?

Top 5%6× random chance

Nuances ML features can't capture

  • Research-stage science, not a workflow tool
    3.3×
  • Founder made real decisions, not just held a title
    2.7×
  • Worked at a breakout company in its early years
    1.9×

+12%vs GPT-6 Sol

$0.00087per prediction

Millisecondsper prediction

Works where ML can'tPitch videosMore use cases
Court litigation Earnings calls Clinical trials Insurance claims and photos Loan applications Policy petitions

Why Think-Reason-Learn

The strengths of both, without their weaknesses.

Machine learning

TRL keeps

  • Speed
  • Low cost
  • Same answer every time
  • Learning from outcomes

TRL fixes

  • Can't read text or video
  • Needs thousands of rows
  • Numbers, not reasons

AccuracyBaseline

Language models

TRL keeps

  • Reads anything
  • Knows the world

TRL fixes

  • Slow and costly per call
  • Can't show how it decided
  • Changes from run to run
  • Doesn't learn from your outcomes

Accuracy+7% over ML

Think-Reason-Learn

Reads like a language model. Runs like machine learning. Shows its reasons.

  • Learns from 50 examples
  • Milliseconds, $0.00087 a prediction
  • Readable rules, questions and policies
  • #1 on VCBench

Accuracy+20% over ML

Inside the harness

Six open-source methods. Each turns a language model's judgment into questions, rules or policies you can read and check.

Frequently asked questions

What is Think-Reason-Learn?

Think-Reason-Learn (TRL) is the decision harness for predictions. It joins machine learning and language models: a language model reads each case and answers plain-language questions, and TRL learns from past outcomes which questions matter and by how much. The result is language-model judgment at machine-learning speed, with the reasons behind every prediction.

How is TRL different from machine learning?

Machine learning needs clean columns and thousands of examples, and it cannot read text, transcripts or video. TRL reads them directly and learns from about 50 examples, while keeping machine learning's speed and consistency. On VCBench it is 20% more accurate than the best machine learning model.

How is TRL different from asking a language model?

A language model is slow and costly per call, can change its answer from run to run, and cannot show how it decided. TRL gives the same answer every time, shows the questions behind each prediction and runs in milliseconds. On VCBench it is 12% more accurate than GPT-6 Sol.

How fast and how cheap is a prediction?

Milliseconds, and about $0.00087 per prediction on VCBench. A language model writes the questions once, during training. At prediction time a small, fast classifier answers them in parallel, so no large model has to reason about each case.

What can TRL predict?

Any outcome where the evidence is text, transcripts, video or images and context matters: which startups break out, how a pitch comes across, court litigation, earnings calls, clinical trials, insurance claims, loan applications and policy petitions.

How do I start?

Choose Add to your agent to set it up in Claude Code, Codex or Cursor, or install it from GitHub with pip install git+https://github.com/vela-research/think-reason-learn.git. TRL is open source under the MIT licence and works with Anthropic, Google, OpenAI and xAI models.

Who builds TRL?

A research lab within Vela Partners and the University of Oxford. Its methods are accepted at IEEE, a NeurIPS'26 workshop and more. Meet the team.

Add TRL to your coding agent

Two steps in the project where you want to build a model. Your agent does the rest.

Open a terminal in your project and start claude.

  1. Install the library. It needs Python 3.13 or newer.
    pip install "git+https://github.com/vela-research/think-reason-learn.git"
  2. Paste this to your agent, then point it at your data.

    Install the think-reason-learn Python package from GitHub with pip install git+https://github.com/vela-research/think-reason-learn.git (Python 3.13 or newer). Read the quick start at https://thinkreasonlearn.com/docs/getting_started/quick_start.html and the API reference at https://thinkreasonlearn.com/docs/, then help me build a model on my data that shows the rules behind each prediction.

You need an API key for the language model you pick, such as OPENAI_API_KEY or ANTHROPIC_API_KEY. See the quick start

Build a model with us

Predict what was never predictable.

1Any decision

2Data

  • PublicCurated and collected
  • VelaProprietary
  • YoursTranscripts, documents, past decisions

Trained with Think-Reason-Learn

3Your model

Will this case settle before trial?

Emails · depositions · hearings

  • Defendant’s emails sound defensive
  • Plaintiff’s story holds under questioning
  • Judge sounds impatient at hearings
Settles78%
  • <$0.001
  • Milliseconds
  • Auditable
  • Yours

Pre-trained model

FounderBrainNew

Hear how you come across. Get coached on what to say next.

Share a pitch or investor call with your agent. FounderBrain compares it with 773 public interviews of founders and CEOs.

Try FounderBrain
  • Free
  • Works in your agent
  • Transcript not stored

Leo MartinThe Candid Advocate

EnthusiasmMore enthusiastic than 94% of founders

MeasuredEnthusiastic

OpennessAbout as personal as most founders

PersonalReserved

CertaintyFirmer claims than 91% of founders

QualifiedFirm

Explains withMore analytical than 74% of founders

DataStories

Talks most like Jensen Huang91% match

Peer-reviewedNeurIPS'26 Workshop