Platform Architecture

The control plane, the hosted cluster, and the artifact path.

The managed platform has two parts. The hosted control plane accepts evaluation requests, stores results, and serves the dashboard and the API. The RLMesh-hosted cluster runs the evaluations. Your organization’s evaluations and probes run on that cluster, so you do not install or operate anything.

The cluster pulls your images from the RLMesh registry and starts a GPU node when a model needs one. A probe or evaluation can wait a few minutes while a node starts. Run pods reach the internet only through a proxy that allows Hugging Face, PyPI and GitHub; see Model Weights and Checkpoints.

From a run to workloads

A run names models and tasks. The platform creates one workload for each model–task pair; a try at that workload is an attempt. It derives the model servers, environment workers, and runtime processes needed to execute them. Use the dashboard to start and inspect a run. Upload a Model or Environment covers the path from a local container to the catalog.

Two paths for results

Live episode results and logs flow from the cluster to the dashboard as the run progresses. They are buffered for at-least-once delivery, so a connection interruption can delay the live view.

Separately, each attempt writes a manifest and artifacts, such as recordings and trajectories, to object storage. The control plane reads the manifest to finalize the workload. If the live view and stored artifacts disagree, the artifacts are authoritative.

Your own cluster

Evaluate on your Kubernetes cluster. Its agent connects outward, starts evaluation pods, and reports status. No kubeconfig or inbound access is needed. See Bring your own cluster.