Tests / latest

Execution consistency

Do browser claims agree with execution? Isolate worker delivery, graphics conversion and codec realization.

Select checks

Worker lifecycleEight 250 ms attempts; separate 1 s Blob, URL and messaging arms.WebGPU consistencyInteger controls, subgroup tuples and supported explicit sizes.WebGL executionSqrt, mipmaps, compressed alpha and sample allocation.Canvas conversion / textFloat16 miniature, alpha, blank, text and fresh / reused controls.Media realization (optional)Owned clips, RTP claims and recorder choices. 60 s total ceiling.

Supplied environment and repetitions

Run labelController condition (supplied)Repetitions in this documentNative environment / changes

Manual start. Default target ~15 s, 30 s total ceiling; media raises it to 60 s. Export cap 512 KiB. Synchronous browser work can overrun a deadline. Keep this tab visible; use fresh processes for independent repetitions.

Loading diagnostic… Reload if this remains.

Results

No capture yet. Nothing is uploaded or saved as a product session.

Full capture JSON

Immutable export

Compare an exported reference

Import after capture. Schema, corpus, known build, order, requested input and observed settings must agree. Development builds with unknown identity are incompatible. Imported data never starts probes.

Reference JSON

No reference imported.

Interpretation and follow-ups

“Relationship violated” means a named control disagreed under its stated preconditions. A failed requested subgroup pipeline alone does not prove false support. Rendering differences are empirical; neither a stock browser nor a CPU reference is a universal pixel oracle.

Worker deadlines retain early censoring and eventual readiness separately. No entry receipt cannot distinguish nonexecution from nondelivery. Controller labels record an external intervention; this page cannot change CDP attachment.

Media queries, configuration admission, decoder output and HTML consumers remain separate. RTP tuple order is ignored for comparison; duplicate tuples remain counted. No authenticity score or detection claim is made.

Modules run sequentially on one cleanup owner. Results appear after cleanup; unresolved cleanup requires reload. Fixed order and repeated runs warm caches. Match actual executable versions, backend, fonts and providers independently before attributing a difference.

Five-second worker diagnostic · Canvas parity and font calibration · Full WebGPU compute signal