STEP 01
WHAT GETS BUILT
Build the environment
The teams who run the work spec the environment. Reward, reset, state, action - written once by the people who use it.
→MODELS · RL ENVIRONMENTS
The environments your team trains in. Reviewed and signed.
The teams who run the work write the environment. One review, one signed approval, one shared registry. Train your model against it, and keep the weights.
CATALOG
22 environments
Approved environments in the shared registry, each with its owner and version attached.
RUNS
48 runs
Training runs executed against approved environments, each logged to the team that ran it.
SIGNED
Approved once
A reviewer signature, a registry version, and a usage log on every environment that ships.
HOW IT WORKS
Three steps. One registry.
Build the environment. Review and approve it once. Share it across every team from one registry.
STEP 02
WHAT GETS APPROVED
Review and approve
Every submission runs one review: static checks, policy gates, and a human signature before any team can train against it.
→STEP 03
WHAT GETS REUSED
Share across teams
Approved environments land in the registry with version, owner, and usage log attached. One source every team trains from.
WHAT COMES OUT
What your team leaves with.
Every approval leaves a record. Every run leaves a log. Every team trains against the same approved environment, with the proof attached.
01
Environment specs
Reward, reset, state, and action - written down by the people who run the work.
↳ ARTIFACT
02
Approval records
Static checks, policy gates, and the reviewer signature that opened the door.
↳ ARTIFACT
03
Version registry
Every approved environment with its version, owner, lineage, and the runs that depend on it.
↳ ARTIFACT
04
Usage logs
Which team ran it, which run consumed it, and what the result was.
↳ ARTIFACT
WHERE IT FITS
In the loop, this is where you test.
Test the run against the environment. Review what broke. Recruit the specialist. Remember the miss. Approve what's right.
01
Test
● YOU ARE HERE
02
Review
03
Recruit
04
Remember
05
Approve
READOUT · DECISION GRAPH
RELATED MODULES
Next to this in Models.
TRAINING
Train against the same environment.
Every team starts from the same approved spec, and keeps the weights they tune on it. No drift between research and rollout review.
See the page →AUTOPILOT
Run the workflow without the toil.
Schedule runs across approved environments. The catalog is the contract.
See the page →FEDERATED LEARNING
Learn together. Keep it separate.
Train across teams on the same approved environment. Share the updates, not the data.
See the page →RL ENVIRONMENTS
Built by the people who use them.
Bring the environment your team already runs. We review it, sign it, and make it shared. You train against it, and keep the weights.