What waiTransport will build.
- WAI capabilities running on GPU and NPU backends where a device has them, and on the CPU where it does not.
- Integer-exact capabilities kept on the CPU paths that reproduce the reference output byte for byte, until an integer GPU or NPU kernel exists and passes the same byte-equality checks. Float paths used only for capabilities whose tier allows them.
- Each backend checked against the same conformance corpora as the reference runtime before it serves a capability.
Open components, each with its WAI label.
- In the standard
Determinism tiers, entropy-consistency and decode-equivalence, typed in code. A replicate means exactly what its tier guarantees.
- In the standard
Fast integer decode kernels that give the same integer for every output.
Notes Timed on one machine. WebAssembly runs the same kernels single-threaded, untimed.
- In the standard
Native and browser builds produce the same output for the integer-exact capabilities.
Notes Checked in the page, per frame, with a 64-bit FNV-1a hash; receipts use BLAKE3.
- In the standard
A payload's exact output size, working memory and compute, read from its header before decoding.
Notes The header bounds the work; a small payload may describe a large output, and a sink applies its own size policy before decoding.
- In the standard
Byte equality of the integer decode paths, checked in CI on each platform the WAI status page lists.
- In the standard
Energy meters that yield a hardware class only when they read hardware.
Notes A meter that cannot measure returns no figure; it never falls back to a fabricated one.
Planned, in the standard and in the service.
- Roadmap· waiTransport
GPU and NPU backends.
- Roadmap· waiTransport
Naming the device and the arithmetic in a receipt. The session claim names the decoder (a built-in decoder and its version, or a module by digest), the decode path within it and the pinned parameter set's tier; it has no device or precision field. Adding one is new work that is neither in the standard nor on the WAI roadmap.
- Roadmap· WAI
A host that runs signed, revocable modules.
What this page does not claim.
- No GPU or NPU backend exists today, in the WAI reference implementation or in waiTransport. A sink may register its own decoder for a capability, as the standard allows.
- At entropy-consistency, the stages after the entropy decode may differ by device. Byte-for-byte output is promised only at decode-equivalence.
- The integer decode kernels are CPU code and cannot use the float GPU and NPU paths; no integer GPU or NPU kernel exists.
- A runtime on other hardware is not a different format. What it decodes is a WAI file that any conforming runtime reads.
- No speed or energy figure for a GPU or NPU backend is stated until it is measured under the published method.
- The reference runtime stays open and free. waiTransport never charges for it.