docs(site): Q&A — can Traverse run models on a GPU or Neural Engine?

public/llms.txt

@@ -79,6 +79,7 @@ Current packages: crates.io Traverse at 0.14.0; npm `traverse-embedder-web@0.14.
 - [How do hosts trust signed models?](https://traverse-framework.com/questions/how-do-hosts-trust-signed-models.html): discover → execute → trace; one shared `runtime.wasm`; host-owned trust roots gate `register_package`; pin by digest + `model.sig.json`; test-only digits-mlp key; published exact-ref execute is native+web+Swift ExactModelHost (`swift-host-v0.14.0-4`); Kotlin and .NET ExactModelHost are on main, not in the published 0.14.0 packages (#1651) — no claimed ONNX generic runner.
 - [Which hosts run signed exact-ref models?](https://traverse-framework.com/questions/which-hosts-run-signed-exact-ref-models.html) · [Is production model signing ready?](https://traverse-framework.com/questions/is-production-model-signing-ready.html) · [Can Kotlin or .NET run exact-ref yet?](https://traverse-framework.com/questions/can-kotlin-or-dotnet-embedders-run-exact-ref-yet.html): discover → execute → trace; one shared `runtime.wasm`; published exact-ref is native + web 0.14.0 + Swift `swift-host-v0.14.0-4`; Kotlin and .NET ExactModelHost are on main (#1580, #1602 closed) and not in Maven/NuGet 0.14.0 (#1651); test-only key ≠ prod (#1567).
 - [What is ExactModelHost?](https://traverse-framework.com/questions/what-is-exact-model-host.html): host-side connector for governed exact-ref model packages (Spec 138). Register a signed package under host-owned trust roots, then execute under discover → execute → trace. Not a second runtime. Published: native connector + web `ExactModelBrowserHost` in 0.14.0 + Swift `ExactModelHost` on `swift-host-v0.14.0-4`. Kotlin/.NET ExactModelHost code is on main, not in Maven/NuGet 0.14.0 (#1651). Test-only digits-mlp key; prod signing #1567. Not catalog `model.execute`. Not Hugging Face.
+- [Can Traverse run models on a GPU or the Apple Neural Engine?](https://traverse-framework.com/questions/can-traverse-run-models-on-a-gpu-or-neural-engine.html): not in a published release; v0.14.0 exact-ref models run on `wasm-cpu` in the one shared runtime. Unreleased on main (Decision 112, traverse#1667): host policy lists ordered placements (default `wasm-cpu`), on-device self-check against the variant's conformance vector, memory ceiling, fail-closed `placement_unavailable`, placement evidence in every result. No real accelerator adapter yet: Core ML conversion spike #1663, iPhone evidence #1630; no web/Android/desktop GPU adapter.
 - [What is digits-mlp?](https://traverse-framework.com/questions/what-is-digits-mlp.html): first trained exact-ref package in v0.14.0 (UCI digits MLP, 96.10% held-out, bit-identical); test-only signing; not a general LLM.
 - [Does the website use the latest traverse-embedder-web?](https://traverse-framework.com/questions/does-the-website-use-the-latest-traverse-embedder.html): site pins `^0.14.0` (matching npm) and `/discover` admits its signed demo fixture through `registerPackage`; the site can lag future releases, so product apps pin published packages, not the website demo.
 - [How does Traverse complement Hugging Face?](https://traverse-framework.com/questions/how-does-traverse-complement-hugging-face.html): Hub = provenance; Traverse = pinned signed client-first capability; not Transformers.js/Hub replacement; no defer promise until a second executor exists.

src/pages/questions/can-traverse-run-models-on-a-gpu-or-neural-engine.astro

@@ -0,0 +1,52 @@
+---
+import QuestionLayout from '@layouts/QuestionLayout.astro';
+
+const jsonLd = JSON.stringify({
+  '@context': 'https://schema.org',
+  '@type': 'FAQPage',
+  mainEntity: [{
+    '@type': 'Question',
+    name: 'Can Traverse run models on a GPU or the Apple Neural Engine?',
+    acceptedAnswer: {
+      '@type': 'Answer',
+      text: 'Not in a published release. In Traverse v0.14.0, governed exact-ref models run on the wasm-cpu path inside the one shared runtime. On traverse main (unreleased), the runtime now has a generic native accelerator adapter framework (Decision 112, Spec 138 FR-060 to FR-065): the host policy lists ordered placements, default wasm-cpu; a placement is chosen only if the host supports it, the signed package carries a variant for it, the variant passes an on-device self-check against its conformance vector, and it fits the host memory ceiling. If nothing qualifies, the call fails closed with placement_unavailable before the input is consumed, and the result records which placement ran. The real Core ML adapter is not landed yet: the conversion spike is #1663 and the iPhone evidence is #1630.',
+    },
+  }],
+});
+
+const relatedLinks = [
+  { href: '/questions/what-is-exact-model-host.html', label: 'What is ExactModelHost?' },
+  { href: '/questions/what-is-exact-ref-model-execution.html', label: 'What is exact-ref model execution?' },
+  { href: '/questions/which-hosts-run-signed-exact-ref-models.html', label: 'Which hosts run signed exact-ref models?' },
+  { href: '/questions/what-happens-when-a-capability-runs-out-of-fuel-or-memory.html', label: 'What happens when a capability runs out of fuel or memory?' },
+  { href: '/questions/is-there-a-performance-cost-to-running-the-same-binary-on-every-os.html', label: 'Is there a performance cost to running the same binary on every OS?' },
+];
+---
+<QuestionLayout
+  title="Can Traverse run models on a GPU or the Apple Neural Engine?"
+  description="Not in a published release. v0.14.0 runs governed exact-ref models on wasm-cpu. Traverse main has an unreleased accelerator adapter framework: policy placements, on-device self-check, fail-closed fallback, placement evidence. Core ML adapter not landed yet."
+  canonical="https://traverse-framework.com/questions/can-traverse-run-models-on-a-gpu-or-neural-engine.html"
+  category="Runtime and architecture"
+  relatedLinks={relatedLinks}
+  jsonLd={jsonLd}
+>
+  <p><strong>Short answer:</strong> not in a published release. In Traverse v0.14.0, governed exact-ref models run on the <code>wasm-cpu</code> path inside the one shared runtime. Hardware acceleration is being built on traverse main, and the part that has landed is the policy and evidence layer, not a working GPU or Neural Engine adapter.</p>
+
+  <h2>What is on main today (unreleased)</h2>
+  <p>As of October 13, 2026, <code>traverse-runtime</code> on main has a generic native accelerator adapter framework (Decision 112, Spec 138 FR-060 to FR-065, <a href="https://github.com/traverse-framework/traverse/pull/1667">traverse#1667</a>). It is not in any published crate or package yet.</p>
+  <ul>
+    <li><strong>The host chooses, in policy.</strong> The execution policy lists placements in order. An empty list means <code>wasm-cpu</code> only. An unknown or repeated placement id is <code>policy_denied</code>.</li>
+    <li><strong>A placement has to earn its turn.</strong> The runtime takes the first listed placement that the host supports, that the signed package carries a variant for, that passed its device self-check, that fits the host memory ceiling, and that meets <code>require_fuel_metering</code> when the host asks for it.</li>
+    <li><strong>The self-check is real.</strong> The variant's conformance vector runs on the device and is compared under the variant's declared tolerance. The result is cached per variant digest, adapter, platform version, and compute units, so an OS update re-runs it. A failed check makes that variant unavailable.</li>
+    <li><strong>It fails closed.</strong> If nothing qualifies, the call fails with <code>placement_unavailable</code> before the input is consumed. A late adapter result after cancel or deadline is discarded.</li>
+    <li><strong>The trace says where it ran.</strong> Every result carries placement evidence: the placement, the compute units, and whether fuel metering applied.</li>
+  </ul>
+
+  <h2>What is not done yet</h2>
+  <p>No real accelerator adapter has landed. The first one planned is Core ML on Apple devices, called from Rust. An earlier converter wrapped ONNX in a way Core ML could only run on the CPU, so it is being replaced through a spike-gated ONNX to Core ML route (<a href="https://github.com/traverse-framework/traverse/issues/1663">#1663</a>), with on-device iPhone timing evidence tracked in <a href="https://github.com/traverse-framework/traverse/issues/1630">#1630</a>. There is no GPU adapter for the web, Android, or desktop, and no date for one.</p>
+
+  <h2>Why it is built this way</h2>
+  <p>An accelerator is an implementation detail of the host, not a second runtime. The contract, the signed package, and <strong>discover → execute → trace</strong> stay the same. The adapter only runs a variant the package already signed, only after it matched the <code>wasm-cpu</code> reference on that device, and the trace records which path produced the answer. An agent or app can ask for speed, but the runtime decides whether the faster path is trustworthy on this hardware.</p>
+
+  <p>This is a different thing from capability <a href="/questions/how-do-placement-targets-work.html">placement targets</a> (<code>local</code>, <code>browser</code>, and so on), which decide where a capability can be placed. Model placements decide which compute path runs a signed model inside a host.</p>
+</QuestionLayout>

src/pages/questions/index.astro

@@ -58,6 +58,7 @@ const groups = [
       ['what-is-one-shared-runtime-wasm.html', 'What is one shared runtime.wasm in Traverse?'],
       ['does-the-rust-embedder-use-the-same-runtime-wasm.html', 'Does the Rust embedder use the same runtime.wasm?'],
       ['what-is-exact-model-host.html', 'What is ExactModelHost?'],
+      ['can-traverse-run-models-on-a-gpu-or-neural-engine.html', 'Can Traverse run models on a GPU or the Apple Neural Engine?'],
       ['how-does-traverse-execute-wasm.html', 'How does Traverse execute WASM?'],
       ['can-traverse-run-without-wasm.html', 'Can Traverse run without WASM?'],
       ['what-is-the-capability-registry.html', 'What is the capability registry?'],

src/pages/questions/what-is-exact-model-host.astro

@@ -20,6 +20,7 @@ const relatedLinks = [
   { href: '/questions/how-do-hosts-trust-signed-models.html', label: 'How do hosts trust signed models?' },
   { href: '/questions/is-production-model-signing-ready.html', label: 'Is production model signing ready?' },
   { href: '/questions/can-kotlin-or-dotnet-embedders-run-exact-ref-yet.html', label: 'Can Kotlin or .NET run exact-ref yet?' },
+  { href: '/questions/can-traverse-run-models-on-a-gpu-or-neural-engine.html', label: 'Can Traverse run models on a GPU or Neural Engine?' },
 ];
 ---
 <QuestionLayout