Skip to content
ReificBacked by Y Combinator

Superhuman reasoning over time.

RL environments for agents: real records that arrive on a clock, checked against what actually happened.

We use your email only to send the samples. Privacy policy

Sell your dataTalk to a researcher

Environments

  1. 1

    Personal agents

    Timestep-Assist. An assistant on someone’s phone. Health, home and messages arrive while it works.

  2. 2

    Industry

    Timestep-Ops. Wind farms, offshore wells, refineries and factories, with months of operating records.

  3. 3

    Science and compute

    Timestep-Lab. Bioreactors, GPU clusters and storage fleets, from the lab to the data centre.

Leaderboard

The mean score of each model on each family, with its 95% interval. Models that train on our environments improve.

Personal agents

Timestep-Assist

  1. Gemini 4 Argon31.6%±5.0%
  2. Claude Opus 5.527.0%±3.6%
  3. Claude Sonnet 5.517.4%±3.7%
All 20 models

Industry

Timestep-Ops

  1. Claude Opus 5.543.1%±5.4%
  2. Gemini 4 Argon35.0%±5.6%
  3. Claude Sonnet 5.534.1%±5.8%
All 20 models

Science and compute

Timestep-Lab

  1. Claude Opus 5.529.7%±14.6%
  2. GPT-6 Astra22.2%±11.4%
  3. GPT-6.1 Sol16.7%±5.7%
All 20 models

Version 17, updated 4 November 2026. Read the methodology.

Research

All

Questions

What are Reific's RL environments?

Tasks to train and test AI agents on real records. Data arrive on a clock while the agent works, and each reward is checked against the data. There are three families: personal agents, industry, and science and compute.

Which model scores highest?

Gemini 4 Argon leads Personal agents with 32%. Claude Opus 5.5 leads Industry with 43% and Science and compute with 30%. The leaderboard above shows the top models of each family.

How are the scores graded?

Each answer key is checked against the records. Numeric checks need no judge, and two judges from different labs grade the rest. A score is a model's mean over the environments of a family, with its 95% interval.

Can we license the environments for RL training?

Yes. We license environments for RL training, and models that train on them improve. Samples are available the same day.

Do you build environments to order?

Yes. We build new environments for a domain, a data shape or a skill that you want to train or measure.

Can we sell or license our data to you?

Yes. We buy licenses to data for AI training in every field, niche ones included. Partner with us.

Try them on your models.

We use your email only to send the samples. Privacy policy

Talk to a researcher

Partner with us.

AI has nearly used up the public web. If you hold data or code no one else has, see what it's worth.

See what your data is worth