Custom training environments for your AI agents.

We build a sandboxed environment of your operations domain and a replay-backed eval suite on top of ConeReplay, so you can measure whether a policy or agent change actually improves outcomes - before it touches production.

What a pilot delivers.

1. 1

   A sandboxed environment of your domain

   Versioned observation/action schema, seeded fixtures, bounded episodes. No production credentials, no network, no clock - the same seed always produces the same world.

2. 2

   Gold and adversarial trajectories

   A gold trajectory that scores full reward, plus adversarial and invalid trajectories with known outcomes - so the reward function is tested, not assumed.

3. 3

   A replay-backed eval suite

   One command runs your policy set over fixed seeds. Every episode is captured as a ConeReplay trace; the report carries reward distributions, failure reasons, and trace-level divergence between policies. Two runs are byte-identical.

Candidate domains.

Status labels are literal: “runnable today” means the code is in the public repo with tests; “spec stage” means design only. Your domain does not need to be on this list - scoping a new one is the pilot.

How a pilot runs.

Discovery call, then a fixed-scope statement of work, then the build, then an evidence handoff: the environment, the eval suite, and the traces. We quote only after discovery - there is no standard price, and we do not claim training gains we have not measured on your workload.

• You run a post-training, eval, or platform team with an agent in or near production.
• You have recorded traces, or a workflow you can describe concretely enough to fixture.
• You need reproducible evidence a policy change helps - not a benchmark from someone else's stack.