⌘K
ENES
GitHub
phase~1 min readUpdated 2026-07-16#phase#benchmarking#security#verification

Always Active — Performance, Security & Verification

This is not a final cleanup phase. It starts with the first WAT module and remains active through browser runtimes, accelerators, offline models, WASI components, and sandboxed tools.

Branch

  • Wasm AI Performance, Security & Craftsmanship — benchmark protocol, cold/warm behavior, latency distributions, throughput, memory, copies, binary/model size, golden outputs, compatibility, integrity, and release evidence.

Permanent obligations

  • Pin source revision, toolchain, dependencies, model, conversion, runtime, backend, browser, operating system, and hardware.
  • Keep deterministic fixtures and golden outputs under version control with documented tolerances.
  • Separate download, compilation, instantiation, model load, first inference, warm inference, and postprocessing.
  • Preserve raw samples and report p50/p95 rather than a context-free average.
  • Record peak memory, tensor-copy costs, Wasm binary size, model size, and storage use.
  • Test unsupported features, forced fallbacks, corrupt artifacts, denied capabilities, traps, and cancellation.
  • Apply CSP, isolation requirements, artifact checksums, provenance, and reproducible-build evidence where relevant.
  • Publish support as a dated observed matrix, never as an unqualified browser claim.

Workbench report contract

Every v0–v8 milestone emits a machine-readable report and a human summary containing runtime, backend, browser, operating system, hardware, model version and size, Wasm binary size, cold start, warm inference p50/p95, peak memory, output checksum, numerical difference, and observed fallbacks. Raw samples and method metadata remain available behind the summary.

Exit criteria for every milestone

  • The artifact builds or runs from documented commands.
  • The output matches a golden reference or explains a reviewed difference.
  • Performance claims include environment, method, distribution, and raw evidence.
  • Backend and fallbacks are visible in both UI and report.
  • Security boundaries and unavailable capabilities have negative tests.
  • Another run can identify every input needed to reproduce or investigate the result.

Use Reference Registry to normalize changing specifications, runtime versions, model artifacts, compatibility observations, and benchmark evidence.

Related notes

Wasm AIPhase 00 — OrientationEstablish prerequisites, platform boundaries, reading routes, compatibility honesty, and the verification contract for WebAssembly AI Atlas.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.Wasm AIPhase 03 — From Models to Browser InferenceConvert real model artifacts, compare browser runtimes, and make preprocessing and postprocessing part of the verified inference contract.