Explanations

Background on the T&E concepts CheckMAITE measures, and on how CheckMAITE works and why it is built that way. These pages don't walk through a task; for that, see the tutorials and how-to guides.

• Saliency and explainability

  What saliency maps show, white-box versus black-box explanation, and how RISE and D-RISE work.

• Distribution shift: drift and OOD detection

  How operational data departs from training data, and how drift and out-of-distribution detection find it.

• Key concepts: capabilities, runs, and caches

  The core abstractions behind how evaluations are defined, executed, and cached.

• Job submission overview

  Why CheckMAITE has non-blocking job handles and pluggable execution backends.

• Job protocol and lifecycle

  The shared job handle contract, lifecycle states, reference-first results, and error semantics.

• Ray job backend

  Why Ray, the registry-backed execution model, status mapping, and design trade-offs.

• Ray simple job backend

  How the process-local Ray backend runs and tracks tasks from a single driver.

• Worker environments

  How container images, worker setup, and runtime_env overlays fit together.

• Kubernetes and KubeRay

  KubeRay placement, detached actors, autoscaling, and durability boundaries.

• Analytics store in distributed execution

  Why durable result writes are more subtle in distributed execution, and what job submission expects from the store.