Environments for frontier models.
Reinforcement learning environments that train frontier models to expert level, graded against ground truth, grounded in real production work. Each plate is a measurement.
Idler / Vol.01
Environments verifiable tasks · dense reward · real rollouts, as measured forms
Method from a problem space to evaluating environments
Domains where we collaborate, in priority order
Why Idler the neutral record
01GroundedEnvironments from real production work, not invented.Real
02NeutralA record measured the same way for every lab.Record
03BroadAcross the problem space and its sub-spaces.Scope
About mission and the neutral record
01MissionTrain frontier models on environments built from real problem spaces, graded against ground truth.Thesis
02Neutral recordA corpus measured the same way for every lab.Record
03TeamA small team, working quietly with frontier labs.Studio
Blog research notes and method write-ups
N1Shelf LifeRepresenting a problem space in thirty pages.Note
N2Environments under RLWhat our environments do to models when applied with RL.Study
N3Dense rewardWhy step-by-step grading beats pass or fail.Note
Careers open roles
01CollaboratorsRun this process with new people. Priority: Safety, Defense, Science, Commerce.Open
02Environment engineeringBuild and scale environments across problem spaces.Role
Contact request access and partnerships
01Request accessSee the environments and what they measure.Access
02PartnershipsRun the process together on a problem space.Partner
Tell us the capability your models miss. We will measure an environment for it.
Request access