⌘K
ENES
GitHub
phase~1 min readUpdated 2026-07-16#phase#orientation

Phase 00 — Orientation

Orientation prevents the Atlas from becoming a collection of unrelated APIs. Before writing kernels or loading models, identify the host, guest, memory boundary, model artifact, runtime, backend, device, capability policy, and evidence required for a claim.

Orientation pages

  • Start Here defines prerequisites, platform roles, and reading routes.
  • Must Know fixes the distinctions between Wasm, WebGPU, WebNN, WASI-NN, runtimes, and backends.
  • WebAssembly AI Index maps every phase and branch.
  • Reference Registry records versions, proposal status, support, artifacts, and benchmark evidence.

Questions to settle

  • Which prior systems and AI concepts are prerequisites for the path you chose?
  • Which tasks run in JavaScript, Wasm CPU, WebGPU, WebNN, or a WASI host?
  • Which browser or runtime features are required, optional, experimental, or unavailable?
  • What is the smallest executable artifact that can prove the current layer?
  • What golden output, tolerance, and benchmark context will verify it?

Phase deliverable

Produce a written execution contract for the first Workbench milestone: target browser, toolchain, host/guest boundary, required features, fallback, fixture, expected output, and report fields. No performance target is valid until this context exists.

Exit criteria

  • You can explain why WebAssembly is not WebGPU and why WebNN is not WASI-NN.
  • You can name the source Atlas to consult instead of duplicating missing prerequisites.
  • You have a dated compatibility assumption and a visible fallback policy.
  • You know how the first output will be compared with a reference.

Continue with Phase 01: WebAssembly Foundations.

Related notes

Wasm AIAlways Active — Performance, Security & VerificationApply benchmarking, numerical regression, compatibility, security, integrity, and reproducibility to every WasmAI Workbench milestone.Wasm AIMust KnowThe minimum distinctions and evidence rules needed to reason correctly about portable, local, accelerated, and sandboxed AI.Wasm AIPhase 01 — WebAssembly FoundationsLearn modules, WAT, validation, toolchains, linear memory, ABI, and host interop before building inference kernels.Wasm AIPhase 02 — Inference from First PrinciplesBuild tensor operations and a tiny forward pass, then measure scalar, SIMD, and threaded Wasm without surrendering numerical correctness.