RL Environments — Glasa Lab

If you can't check the reward,
you can't trust the policy.

Run millions of episodes.Same seed, same state, same grade.

InterfaceGym-style reset() / step()
ActionsTool calls, browser, GUI, code
RewardsProgrammatic, rubric or hybrid
PackagingOCI images, scale on your cluster

Environment catalog

Enterprise · long-horizon

Procurement & ERP

Raise POs, match invoices and chase approvals across a seeded ERP, inbox and chat. Graded on policy, budget and timing.

Enterprise · tool use

Support triage

Route, resolve and escalate tickets against real policy docs and SLAs. Graded with code checks plus expert rubrics.

Enterprise · exact match

Finance close

Reconcile ledgers and build month-end packs in a spreadsheet sandbox. Every number is checked.

CAD · geometric reward

Parametric CAD edit

Change features to hit a spec. Graded on geometry, constraints, mass and rebuild success.

CAD · physics check

Assembly & mates

Assemble parts with correct mates and clearances. Interference and motion checks score the result.

CAD · verifiable

Text-to-CAD

Turn a written requirement into a valid, manufacturable model. Scored against hidden reference geometry.

Sample task

What your trainer sees.

task.pyimport glasa_envs

env = glasa_envs.make("procurement-erp-v3", seed=48112)
obs = env.reset()

while not done:
    action = policy(obs)          # tool call or GUI action
    obs, reward, done, info = env.step(action)

# info["grader"] →
#   po_valid      True
#   under_budget  True
#   policy_ok     True
#   eta_on_time   True
#   reward        1.00

Graders

Hard to hack, easy to read

Each grader checks end state, not the transcript. We red-team every environment for reward hacks before release and publish the audit.

DeterminismSeeded, reproducible
ThroughputParallel, stateless workers
SplitsTrain / held-out, refreshed monthly

Can we run environments on our own infrastructure?

Yes. Environments ship as container images and run entirely on your cluster.

Do you build custom environments?

Yes. Tell us the capability you're training and we'll scope the app replica, the task set and the grader.

How do you prevent reward hacking?

Graders check verified end state, and every environment goes through adversarial testing with strong policies before release.

Need a custom environment?

Scope a Build

If you can't check the reward,you can't trust the policy.