The physical world is missing
from the training set.
Manuals capture how machines should work. The knowledge of how they actually fail lives in the hands of people who build, operate, and repair them. We turn that tacit expertise into interactive, physics-grounded tasks with outcomes models cannot talk their way around.
One environment. A general method.
no-start-env is our first public proof: an agent gathers evidence, builds a causal picture, takes action, and is graded by what the simulated system actually does. That pattern extends from fault isolation to robot recovery, maintenance, and operation.
A physical-reasoning gap you can measure.
Our first benchmark uses 5 vehicle electrical scenarios × 5 epochs with outcome-based grading. Showing 7 of 9 models; full table in the repo.
Training gyms and test rigs for physical intelligence.
Physics, not scripted answers
Agents interact with a live system model, not a tree of hand-authored responses. State, measurements, and consequences are computed from the governing physics, so each task can vary without breaking the rules of the machine.
Outcomes, not plausible text
Graders reward correct causal understanding, disciplined action, and efficient recovery against an expert baseline. The environment state decides whether an intervention worked; confident explanations and brute-force guessing do not.
Expert verification is the moat
Models can generate simulations they cannot warrant. Every scenario is played and signed off by someone who understands the real system. In no-start-env, that review caught four plausible physics bugs that had passed the automated suite.
From model capability to machine reliability.
Built for frontier labs training general agents, robotics teams testing deployment readiness, and operators turning hard-won machine knowledge into repeatable practice. We measure whether a model can observe, localize, decide, and recover under physical constraints.
Plans that break the physics.
Coaching dependence.
Inspect our first environment.
no-start-env is open source and pip-installable. See how a physical system model, instrumented action space, outcome-based grader, and expert-verified invariants fit together — then run the benchmark on any Inspect-supported model.
Build the missing
training ground.
We work with AI labs, robotics teams, and industrial operators to turn real machine behavior and hard-won domain knowledge into interactive training environments and rigorous evals. If your models must reason before they act, we'd like to talk.
