Skip to main content
Train is the dream-training API. Where the Play API creates and embeds playable worlds, Train uses a world model as a training environment: your policy learns inside the dream, then gets examined in a frozen physics exam it never trains in.

Train in the dream

CMA-ES evolves a compact controller inside the doom world model, the same public checkpoint as alakazamworld/doom-dungeon-hg (byte-identical, sha256-verified).

Certify in the oracle

A frozen Webots exam (two physical worlds, fixed contract, anti-exploit control arms) is the sole scoreboard. Dream fitness is never a capability claim.

​
The pipeline

One POST /jobs runs both stages and leaves a machine-readable job.json:
Jobs are asynchronous, kill-tolerant, and append-only. Poll GET /jobs/{job_id} for live progress; never block on the POST.

​
Lineage is an explicit input

Warm-start any job from a prior champion with train.parent (an inline genome or a prior job_id). Lineage is recorded in the artifacts, so chains are reproducible with no ambient state. Warm-started jobs reproduce the historical best-known profile: contact-safe, 0 contacts in 40/40 exam episodes.

​
Two ways to train

Server jobs (this API)

Submit a spec, get a reproducible, certified result. Best for lineage chains and anything you want the oracle’s verdict on.

Local gym (your hardware)

The world-model weights are public. Run the dream locally (WebGPU or ONNX Runtime) at your own speed, then submit champions to the exam. See Local gym.

​
What’s live vs. roadmap

The serving VM sleeps when idle; the first request wakes it. During the boot window (4 to 8 min) requests return 503 {"status": "waking"} with a phase field; retry the same call until it succeeds. Measured wall-times on the serving VM: smoke job 12–17 min, real-scale job 3–4 h, exam alone 4–6 min (the VM’s CPU generation varies per boot).