Get in touch

ABOUT US

Why we started Research Lab

Intelligent systems are becoming powerful faster than they're becoming trustworthy. We started Research Lab to help close that gap.

Our Story

Built by people who got tired of asking systems to be trusted

Everywhere we looked, AI was giving confident answers that nobody could check. A 2024 Stanford study found leading legal research tools wrong or unsupported on roughly 17 to 33% of test questions. Forecasts promised 80% certainty and delivered closer to 60%. Engineers redesigned parts that already existed because no search could find them.

We started with a simple rule: an AI system should show its evidence. That means testing every model on the customer's own data, deciding how we'll judge an experiment before we run it, and building assistants that point to their sources and admit when they don't know.

Today we're a small lab working across learning agents, trustworthy forecasting, and AI search for law and engineering. We publish our setbacks alongside our results, including the dry run where our own agent market lost by 11 points.

Image

What we believe

Our research philosophy

A handful of principles that shape which projects we take on, and how we run them.

Evidence, not claims

Every figure we state should have a test behind it. When the evidence is missing, we say so plainly.

Decide the test first

We fix our hypotheses and settings before running an experiment, so a result can't be tuned into a win after the fact.

Test on real data

We test on each customer's own data, a firm's own questions and realistic part libraries, never on toy examples alone.

Know when to say
"I don't know"

A good system says clearly when nothing suitable exists, and leaves the judgment calls to people.

MILESTONES

Where our work stands

What's done, what's under way, and what comes next.

  1. DONE

    Forecasting foundations

    A survey of forecasting methods and a catalogue of 287 datasets across
five sectors, from retail to financial services.

  2. DONE

    First forecasting benchmark

    Twelve methods tested on ten public datasets. It showed that no single model wins everywhere, for about $2.70 of GPU time.

  3. DONE

    Agent market simulator

    A tested simulation of our agent credit market, with 31 checks on credits, auctions and memory limits. It caught a real design flaw before any live run.

  4. DONE

    Experiment pre-registered

    The full design for our agent market study is fixed in advance: 5 comparison arms, 10 seeds and 30 rounds, with strict go/no-go checks.

  5. DONE

    Legal AI assessment

    A practical map of seven kinds of legal work, rated by importance and by how far AI has actually solved them.

  6. IN PROGRESS

    MECH-RAG research plan

    Four connected research topics for AI search across CAD models and
technical drawings, plus realistic benchmarks to measure it.

  7. NEXT

    Live experiments and customer testing

    Running the agent market with real AI agents, and building UPPi's per-customer test before it serves any real forecasts.

4
Research programmes
287
Datasets catalogued
12
Forecasting methods
11
Research notes published

The people behind it

Meet the team

Meet The Full Team

Interested in collaborating?

Whether it's a research partnership, a visiting researcher position, or real data you'd like tested,
such as part libraries, legal archives or forecasting problems, we'd like to hear from you.

Get In Touch