Live environments, not multiple-choice trivia
The agent takes actions, breaks things, reads error tracebacks, and fixes its own mistakes on a fresh sandbox every turn.
Commercial OpenEnv marketplace
Flight simulators for AI agents.
An agent can't master a real engineering job by reading static textbooks. Private training gyms give the model a sandbox with real tools (terminals, CAD, compilers), throw tricky edge cases at it, and score every attempt with an automated judge.
Real tool calls, not prompt tricks
Automated binary grading
Zero data leakage
OpenEnv standard API
Why g factor
Why own the weights instead of renting an API?
Zero tolerance for polite fluff
The model cannot earn points by writing a smooth explanation. The code has to compile, the circuit has to regulate voltage, and the robot has to not crash.
Your secret sauce stays inside
Run state, proprietary simulators, test fixtures, and customer data never leave your private network. Only numeric reward signals cross the wire.
Capabilities
Specialized environments backed by executable verifiers.
Kernel / Triton optimization
Optimize PyTorch operators with Triton or CUDA while hidden numerical cases and stable GPU timing gate the speed reward.
- Numerical equivalence across hidden shapes
- Stable performance measurement
- Correctness before speed reward
Digital design and formal verification
Train agents to implement synthesizable Verilog RTL and prove held-out Lean 4 theorems against real compiler, simulation, formal, timing, and axiom checks.
- Icarus, Verilator, Yosys/SKY130, and formal equivalence
- Lean 4 elaboration in frozen mathlib context
- Hidden cases and machine-audited correctness
Physics and silicon engineering
Optimize coupled Modelica systems, OpenFOAM/SU2 airfoils, and SKY130 analog circuits against hidden operating conditions and native simulation evidence.
- Solver convergence and conservation balances
- CFD mesh, residual, lift, and drag checks
- PVT, noise, DRC/LVS/PEX, and post-layout replay
CAD, manufacturing, and 3D workflows
Create mechanical solids, FreeCAD CAM plans, Blender scenes, and structured DXF drawings that are evaluated as native artifacts rather than screenshots.
- CadQuery/OpenCascade and STEP/STL geometry
- Collision-safe CNC operations and reparsed G-code
- Blender scene state and entity-level DXF checks
Quantitative finance research
Train agents to express original cross-sectional alpha hypotheses in a restricted formula DSL and improve them against frozen out-of-sample evidence.
- Purged cross-validation and next-period rank IC
- Trading-cost, turnover, and backtest-overfit controls
- Novelty, complexity, and adversarial regime stress
Game and rendering engineering
Write portable GLSL 330 shaders that generalize across hidden scenes and stay inside a measured frame-time budget.
- Compiler and image-error feedback
- Hidden uniforms, transforms, and lighting
- Hard 16 ms frame gate
Before you start
Common questions
How is a training environment different from a dataset?
A dataset is like reading a textbook: it shows examples of completed work. A training environment is like a flight simulator: the model sits in the cockpit, hits buttons, makes mistakes, and learns how to recover when things go wrong.
Can I connect my own OpenEnv environment?
Yes. You can bring any proprietary simulator, internal API, or test harness. As long as it can send actions and return observations and a reward score (via OpenEnv), g factor can train an agent against it.
Which engineering workflows do the environments cover?
Our 92-gym catalog covers robotics, chip design (Verilog, analog IC), multiphysics (Modelica, CFD), CAD/CAM, theorem proving (Lean 4), quantitative finance, and systems engineering. You can explore the full list in our catalog.
Private beta
Turn your workflow into a skill your model can own.
Bring a model and a task that matters to your company. We'll tell you honestly if post-training fits.
Request Demo