Type a prompt or choose a sample, set the target budget or compression ratio, and see exactly which tokens are kept (sinks, semantic anchors, and recent tokens) vs those evicted dynamically at runtime.