Backed by Y CombinatorRL environments for reasoning over time.
Reific builds reinforcement learning environments for reasoning over time. In each environment the world keeps moving while the agent works: data keep arriving on a clock, people reply on their own schedule, and every reward is checked against what the data later show.
Environments
Personal agents, Timestep-Assist
An assistant on someone’s phone. Health, home and messages arrive while it works.
Industry, Timestep-Ops
Wind farms, offshore wells, refineries and factories, with months of operating records.
Science and compute, Timestep-Lab
Bioreactors, GPU clusters and storage fleets, from the lab to the data centre.
Scores and method
Leaderboard
Each frontier model's average score on each Timestep track, with its 95% interval.
Methodology
How the environments are run and scored.
Work with Reific
Request samples
AI labs request sample environments with a work email. Samples go out the same day.
Request a model
Search a board for a model. If it is not there, ask us to run it.
Data partners
Organisations license records that change over time, such as health, sensor, lab and market data.
Research
How we protect people's data, in layers
Find-and-replace misses names in nicknames, file names and inflected forms, and a record with no names can still point to one person. We protect people's data in eight layers, and models from two labs attack every release before it ships.
Timing is part of the answer
Static benchmarks wait for all the data before grading an agent. For agents that act while events unfold, that measures the wrong skill, and training on it teaches models to wait.
What frontier models show about reasoning over time
Version 7 of our leaderboard puts 15 models on industrial and scientific systems and 14 on personal agents. No model earns half of the possible reward in any family, and different labs lead different families.
Why we build from real records
Real operating records have gaps, drift, late corrections and sources that disagree. Synthetic data rarely has them. An environment built from real records teaches the work as it really is.
How experts shape our environments
Every environment we ship is checked by people who do the work it is about: clinicians for health tasks, engineers and operators for industrial ones, scientists and site reliability engineers for the lab and the data centre. Their decisions are already in the records we build from, their rules set the thresholds, they review how we grade, and every lesson from their reviews goes into the next environment we build.
A curriculum for reasoning over time
Each of our environments has easier levels and a hard level. The best models clear the easier levels and still fail the hard one. That gradient makes an environment useful for training, not only for evaluation.
Frontier evals for LLM reasoning over time series
Frontier models handle simple questions about time series well. Long histories, decision cutoffs and questions with a trap in them still break them. For frontier evals of LLM reasoning over time series data, reach out to us.
EVEREST: when the rare event is the one that matters
A closer look at rare-event forecasting: what a probability means, how an alert changes a decision, and what the evidence supports.
Every article has a Markdown version: add /markdown to its address. Read every article in Markdown or the structured index.
Page reviewed .