Add WebGPU evidence row action

README.md

@@ -28,7 +28,7 @@ Known testing notes:
 - The app automatically falls back to WASM if hardware WebGPU is unavailable.
 - Initial load can be large because STT, LLM, TTS, and voice embeddings are all browser-side assets.
 - For validation evidence, run **Run benchmark suite**, **Run 3 real-mic series**, then **Download JSON**.
-- The Latency panel shows separate real-mic and hardware-WebGPU validation cards.
+- The Latency panel shows separate real-mic and hardware-WebGPU validation cards. On hardware browsers, **Run WebGPU evidence row** captures a fast row for the downloaded JSON.
 - The exported `hostMetadata.hfSpaceCommit` identifies the exact static Space revision served to the browser.
 - Hosted smoke validation passed on May 28, 2026: TTS, identity, and loopback rows completed with zero benchmark-phase network requests, zero benchmark errors, and zero server-inference suspects.
 - The latest local validation also passed UI smoke, client-side/no-server smoke, synthetic loopback stability, and first-TTS-chunk safety. The remaining external gates are real human microphone rows and a hardware WebGPU row from a browser that exposes a real GPU adapter.

app.js

@@ -10,6 +10,7 @@ const elements = {
   ttsBenchmarkButton: $("ttsBenchmarkButton"),
   loopbackButton: $("loopbackButton"),
   bargeInButton: $("bargeInButton"),
+  gpuBenchmarkButton: $("gpuBenchmarkButton"),
   micBenchmarkButton: $("micBenchmarkButton"),
   micSeriesButton: $("micSeriesButton"),
   copyResultsButton: $("copyResultsButton"),
@@ -88,6 +89,14 @@ const IDENTITY_PRIMER_MESSAGES = [
     role: "assistant",
     content: APP_IDENTITY_ANSWER,
   },
+  {
+    role: "user",
+    content: "Please identify this browser demo.",
+  },
+  {
+    role: "assistant",
+    content: APP_IDENTITY_ANSWER,
+  },
 ];
 const APP_IDENTITY_QUALITY_RULES = [
   { label: "client/browser/local", pattern: /\b(client(?:-side)?|browser|local)\b/i },
@@ -120,9 +129,18 @@ const MIN_VAD_SILENCE_MS = 200;
 const MAX_VAD_SILENCE_MS = 800;
 const DEFAULT_LOOPBACK_SPEED = 1.0;
 const DEFAULT_LOOPBACK_PREROLL_MS = 180;
-const DEFAULT_LOOPBACK_PROMPT = "What application is this?";
+const DEFAULT_LOOPBACK_PROMPT = "Please identify this browser demo.";
 const REAL_MIC_TARGET_RUNS = 3;
 const REAL_MIC_MAX_WER = 0.25;
+const HARDWARE_WEBGPU_EVIDENCE_STACK = Object.freeze({
+  device: "webgpu",
+  llm: "HuggingFaceTB/SmolLM2-135M-Instruct",
+  asr: "onnx-community/moonshine-base-ONNX",
+  voice: "F2",
+  ttsSteps: 2,
+  vadSilenceMs: 480,
+  partialAsr: true,
+});
 const MIC_START_TIMEOUT_MS = 15000;
 const WEBGPU_ADAPTER_TIMEOUT_MS = 5000;
 const DEFAULT_TTS_CHUNKING = Object.freeze({
@@ -469,6 +487,7 @@ function setBenchmarkControlsDisabled(disabled) {
   elements.bargeInButton.disabled = locked;
   elements.micBenchmarkButton.disabled = locked;
   elements.micSeriesButton.disabled = locked;
+  updateGpuEvidenceButton();
 }
 
 function resetMetrics() {
@@ -579,6 +598,21 @@ function updateGpuValidationStatus() {
   elements.gpuValidationStatus.textContent = status;
   elements.gpuValidationDetail.textContent = detail;
   elements.gpuValidationProgressBar.style.width = `${Math.round(Math.max(0, Math.min(1, progress)) * 100)}%`;
+  updateGpuEvidenceButton();
+}
+
+function updateGpuEvidenceButton() {
+  if (!elements.gpuBenchmarkButton) return;
+  const hasHardware = state.webgpuAvailable === true && state.webgpuSoftwareAdapter !== true;
+  const busy =
+    state.modelsLoading ||
+    state.suiteRunning ||
+    state.micSeries.active ||
+    Boolean(state.activeBenchmark);
+  const hardwareRows = evidenceSummaryForRows(state.benchmarkResults).hardwareWebgpu.hardwareRows ?? 0;
+  elements.gpuBenchmarkButton.disabled = busy || !hasHardware;
+  elements.gpuBenchmarkButton.textContent =
+    hardwareRows > 0 ? "Run another WebGPU evidence row" : "Run WebGPU evidence row";
 }
 
 function resetTurnState() {
@@ -1356,7 +1390,11 @@ function updateOutputQuality(benchmark, output) {
 
 function isAppIdentityQuestion(text) {
   const normalized = normalizeForQuality(text);
-  return /^(what's|what is|what app is|what demo is|what application is) this$/.test(normalized);
+  return (
+    /^(what's|what is|what app is|what demo is|what application is) this$/.test(normalized) ||
+    /^(what's|what is) this (app|application|demo)$/.test(normalized) ||
+    /^(please )?identify this (browser )?(demo|app|application)$/.test(normalized)
+  );
 }
 
 function words(text) {
@@ -2302,6 +2340,48 @@ function runBargeInBenchmark() {
   return true;
 }
 
+async function runHardwareWebGpuEvidenceBenchmark() {
+  if (state.activeBenchmark || state.suiteRunning || state.micSeries.active || state.modelsLoading) {
+    logEvent("Finish the current benchmark before collecting WebGPU evidence.");
+    return false;
+  }
+  await supportsWebGPU();
+  if (!state.webgpuAvailable || state.webgpuSoftwareAdapter) {
+    logEvent("Hardware WebGPU evidence requires a browser exposing a real GPU adapter.");
+    updateGpuValidationStatus();
+    return false;
+  }
+  if (!state.modelsLoaded) {
+    applyHardwareWebGpuEvidenceStack();
+    updateRuntimeStatus();
+    logEvent("Loading the default hardware WebGPU evidence stack.");
+    await loadModels({ ttsWarmup: false });
+    if (!state.modelsLoaded) {
+      logEvent("WebGPU evidence stack did not finish loading.");
+      return false;
+    }
+  }
+  const stack = currentBenchmarkStack();
+  if (stack.device !== "webgpu") {
+    logEvent("Unload the current stack, select WebGPU or Auto on a hardware adapter, then rerun WebGPU evidence.");
+    return false;
+  }
+  logEvent("Running hardware WebGPU evidence row.");
+  return runBenchmark();
+}
+
+function applyHardwareWebGpuEvidenceStack() {
+  elements.deviceSelect.value = HARDWARE_WEBGPU_EVIDENCE_STACK.device;
+  elements.llmModelSelect.value = HARDWARE_WEBGPU_EVIDENCE_STACK.llm;
+  elements.asrModelSelect.value = HARDWARE_WEBGPU_EVIDENCE_STACK.asr;
+  elements.voiceSelect.value = HARDWARE_WEBGPU_EVIDENCE_STACK.voice;
+  elements.ttsSteps.value = String(HARDWARE_WEBGPU_EVIDENCE_STACK.ttsSteps);
+  elements.vadSilence.value = String(HARDWARE_WEBGPU_EVIDENCE_STACK.vadSilenceMs);
+  elements.partialToggle.checked = HARDWARE_WEBGPU_EVIDENCE_STACK.partialAsr;
+  updateTtsStepsLabel();
+  updateVadSilenceLabel();
+}
+
 async function runMicBenchmark(options = {}) {
   const series = options?.series === true;
   if (!state.modelsLoaded) return false;
@@ -2690,6 +2770,7 @@ function installAutomationApi() {
     runTts: (options) => runSingleAutomationBenchmark(runTtsBenchmark, options),
     runLoopback: (options) => runSingleAutomationBenchmark(runLoopbackBenchmark, options),
     runBargeIn: (options) => runSingleAutomationBenchmark(runBargeInBenchmark, options),
+    runWebGpuEvidence: (options) => runSingleAutomationBenchmark(runHardwareWebGpuEvidenceBenchmark, options),
     runMic: (options) => runSingleAutomationBenchmark(runMicBenchmark, options),
     runMicSeries: runAutomationMicSeries,
     synthesizeAudio: synthesizeAutomationAudio,
@@ -2734,6 +2815,9 @@ elements.chatBenchmarkButton.addEventListener("click", runChatBenchmark);
 elements.ttsBenchmarkButton.addEventListener("click", runTtsBenchmark);
 elements.loopbackButton.addEventListener("click", runLoopbackBenchmark);
 elements.bargeInButton.addEventListener("click", runBargeInBenchmark);
+elements.gpuBenchmarkButton.addEventListener("click", () => {
+  runHardwareWebGpuEvidenceBenchmark().catch((error) => logEvent(`WebGPU evidence failed: ${error.message}`));
+});
 elements.micBenchmarkButton.addEventListener("click", runMicBenchmark);
 elements.micSeriesButton.addEventListener("click", runMicSeriesBenchmark);
 elements.copyResultsButton.addEventListener("click", () => {

index.html

@@ -190,6 +190,7 @@
             <button id="ttsBenchmarkButton" type="button" disabled>Run TTS benchmark</button>
             <button id="loopbackButton" type="button" disabled>Run voice loopback</button>
             <button id="bargeInButton" type="button" disabled>Run barge-in check</button>
+            <button id="gpuBenchmarkButton" type="button" disabled>Run WebGPU evidence row</button>
             <button id="micBenchmarkButton" type="button" disabled>Benchmark real mic</button>
             <button id="micSeriesButton" type="button" disabled>Run 3 real-mic series</button>
           </section>

styles.css

@@ -426,6 +426,7 @@ h2 {
 #ttsBenchmarkButton,
 #loopbackButton,
 #bargeInButton,
+#gpuBenchmarkButton,
 #micBenchmarkButton,
 #micSeriesButton {
   width: 100%;
@@ -436,6 +437,7 @@ h2 {
 #ttsBenchmarkButton,
 #loopbackButton,
 #bargeInButton,
+#gpuBenchmarkButton,
 #micBenchmarkButton,
 #micSeriesButton {
   margin-top: 8px;