Delegation truth harness

Answers one question: for a given .tflite model, LiteRT.js version and backend, does the model actually run where it claims? This is the one page on the site where numbers are the content. Demo pages are visual-first.

Runtime
Model
Backends

Every timing below is time to usable output — input handoff through the awaited output readback, not raw kernel time. For large-output demos (e.g. 4x upscaling) the readback can dominate the number; see measure.ts.

Backend Delegation load+compile first inf. median p90

Environment

reading…

Captured LiteRT console output

Partial delegation is reported nowhere else. This panel is half the evidence the site rests on.

(nothing yet)