Record Live delta and standby evidence

site/data/live-codex-worker-result.ts

@@ -1,5 +1,5 @@
 export const liveCodexWorkerResult = {
-  generatedAt: '2026-07-13T18:39:45.449Z',
+  generatedAt: '2026-07-13T19:06:57.179Z',
   codexVersion: '0.144.0-alpha.4',
   transport: 'stdio JSONL',
   model: 'GPT-5.6-Sol',
@@ -51,16 +51,16 @@ export const liveCodexWorkerResult = {
   productionBrowser: {
     runs: 1,
     selectedTarget: 'article.offer-card',
-    firstVariantMs: 28823.47,
-    firstVariantP95Ms: 28823.47,
-    secondVariantMs: 72698.01,
-    allVariantsMs: 145234.51,
-    allVariantsP95Ms: 145234.51,
-    firstToSecondMs: 43874.54,
-    secondToAllMs: 72536.5,
-    serverPickupMs: 56.31,
+    firstVariantMs: 34177.39,
+    firstVariantP95Ms: 34177.39,
+    secondVariantMs: 54156.22,
+    allVariantsMs: 122089.41,
+    allVariantsP95Ms: 122089.41,
+    firstToSecondMs: 19978.83,
+    secondToAllMs: 67933.19,
+    serverPickupMs: 54.02,
     sourcePreflightMs: 43,
-    validationMs: 5.67,
+    validationMs: 8.67,
     acceptRuns: 1,
     acceptedVariant: 2,
     acceptToResetMs: 231.1,
@@ -68,6 +68,28 @@ export const liveCodexWorkerResult = {
     acceptToNextFirstVariantMs: 52612.73,
     blockedTailBaselineMs: 68141,
   },
+  fullSourceControl: {
+    runs: 1,
+    selectedTarget: 'article.offer-card',
+    firstVariantMs: 28823.47,
+    secondVariantMs: 72698.01,
+    allVariantsMs: 145234.51,
+    firstToSecondMs: 43874.54,
+  },
+  sourceDeltaExperiment: {
+    baselineFirstToSecondMs: 43874.54,
+    deltaFirstToSecondMs: 19978.83,
+    improvement: 0.5446,
+    variant2Passed: true,
+    status: 'Enabled for Codex source mode',
+  },
+  standbyExperiment: {
+    controlFirstVariantMs: 33283.06,
+    prewarmedFirstVariantMs: 35049.12,
+    controlNextFirstVariantMs: 42969.34,
+    prewarmedNextFirstVariantMs: 50920.73,
+    status: 'Rejected: slower under app-server contention',
+  },
   twoCheckpointControl: {
     runs: 1,
     selectedTarget: 'article.offer-card',
@@ -91,10 +113,11 @@ export const liveCodexWorkerResult = {
     runs: 1,
     variants: 3,
     passed: 1,
-    commandFidelity: 6,
-    brandFidelity: 7.67,
-    renderQuality: 6,
-    taskCompletion: 6,
+    commandFidelity: 6.33,
+    brandFidelity: 7.33,
+    renderQuality: 6.33,
+    taskCompletion: 5.67,
+    variant2Passed: true,
   },
   calibratedHeadingQuality: {
     runs: 3,
@@ -123,5 +146,5 @@ export const liveCodexWorkerResult = {
   },
   result: '4/4 full-task quality gate',
   passed: true,
-  architecture: 'A dedicated persistent Live app-server connection and worker thread with native Impeccable skill, sub-command, project context, and annotation-image inputs. Variants 1 and 2 publish as independent reviewable checkpoints; variant 3 and parameters finish last. Accept interrupts unfinished variants; cleanup stays on the foreground control lane while the next generation starts independently.',
+  architecture: 'A dedicated persistent Live app-server connection and worker thread with native Impeccable skill, sub-command, project context, and annotation-image inputs. Variants 1 and 2 publish as independent reviewable checkpoints; source-mode variant 2 uses a fenced delta while variant 3 and parameters finish last. Accept interrupts unfinished variants; cleanup stays on the foreground control lane while the next generation starts independently.',
 } as const;

site/data/live-harnesses.ts

@@ -56,11 +56,11 @@ export const liveExperiments = [
   },
   {
     rank: 3,
-    title: 'Reveal variants progressively — shipped',
-    evidence: 'Five production Sol/medium runs showed variant 1 at 19.3 s median / 20.4 s p95 and the full set at 52.8 s median / 55.1 s p95.',
-    move: 'Codex publishes a monotonically growing prefix as each variant validates. The browser reveals arrived variants and pending dots immediately; Accept and Discard fence unfinished work.',
-    expected: 'Lets review begin as soon as one credible option exists.',
-    confidence: 'Production browser',
+    title: 'Send variant 2 as a fenced delta — shipped',
+    evidence: 'On the proven article.offer-card target, variant 2 arrived 20.0 s after variant 1 instead of 43.9 s with full-source output, a 54.5% reduction. The rendered judge passed variant 2 at 7/8/7/7.',
+    move: 'Keep the full skill, plan, project context, and source evidence, but return only variant 2 markup and scoped CSS. The supervisor validates its variant identity and merges it transactionally without exposing prior output to rewrites.',
+    expected: 'Makes the second credible option reviewable much sooner without weakening design context.',
+    confidence: 'Production browser + rendered gate',
   },
   {
     rank: 4,
@@ -96,11 +96,11 @@ export const liveExperiments = [
   },
   {
     rank: 8,
-    title: 'Protect rendered quality before trimming context — blocking',
-    evidence: 'Only 1 of 9 rendered /bolder variants passed the multimodal gate. A compact-prompt trial made the progressive tail slower and passed 0/3, so it was reverted.',
-    move: 'Keep the full skill, sub-command, and project context. Improve diversity and fidelity against screenshots before pursuing prompt-size or lower-reasoning shortcuts.',
-    expected: 'Prevents a latency win that produces off-brand or structurally weak frontend work.',
-    confidence: 'Three production runs + rendered judge',
+    title: 'Reject speculative model warmup — decided',
+    evidence: 'A prewarmed spare thread made first review slower (35.0 s vs 33.3 s) and the post-Accept next result slower (50.9 s vs 43.0 s). The app-server workload contended instead of hiding latency.',
+    move: 'Do not spend model work before a real generation request. Keep cold thread rotation for cancellation safety and optimize the requested output itself.',
+    expected: 'Avoids extra tokens, contention, and initialization work that did not improve the user-visible path.',
+    confidence: 'Matched production A/B',
   },
   {
     rank: 9,

site/pages/live-lab/index.astro

@@ -406,7 +406,7 @@ const providerColumns = [
       <div class="live-performance-finding live-finding-card--quiet" role="note">
         <strong>Dedicated Codex worker</strong>
         <p>
-          With the target proven as <code>{liveCodexWorkerResult.productionBrowser.selectedTarget}</code>, variant 1 arrived in {displayMs(liveCodexWorkerResult.productionBrowser.firstVariantMs)}, variant 2 in {displayMs(liveCodexWorkerResult.productionBrowser.secondVariantMs)}, and variant 3 in {displayMs(liveCodexWorkerResult.productionBrowser.allVariantsMs)}. Accepting variant 2 while variant 3 was pending returned the page in {displayMs(liveCodexWorkerResult.productionBrowser.acceptToResetMs)} and dispatched the next Go in {displayMs(liveCodexWorkerResult.productionBrowser.acceptToNextGoDispatchMs)}. In matched full tasks, warm app-server reached usable output in {displayMs(liveCodexWorkerResult.architectureComparison.warmAppServerFirstUsableMs)} versus {displayMs(liveCodexWorkerResult.architectureComparison.directFirstUsableMs)} direct; both passed 4/4.
+          With the target proven as <code>{liveCodexWorkerResult.productionBrowser.selectedTarget}</code>, variant 1 arrived in {displayMs(liveCodexWorkerResult.productionBrowser.firstVariantMs)} and the fenced variant-2 delta followed {displayMs(liveCodexWorkerResult.productionBrowser.firstToSecondMs)} later—{(liveCodexWorkerResult.sourceDeltaExperiment.improvement * 100).toFixed(1)}% faster than full-source output. Variant 2 passed the independent rendered gate. Accepting it while variant 3 was pending returned the page in {displayMs(liveCodexWorkerResult.productionBrowser.acceptToResetMs)} and dispatched the next Go in {displayMs(liveCodexWorkerResult.productionBrowser.acceptToNextGoDispatchMs)}.
         </p>
       </div>
       <p class="table-scroll-hint" aria-hidden="true">Scroll horizontally to compare →</p>
@@ -416,12 +416,20 @@ const providerColumns = [
           <thead><tr><th>Candidate</th><th>Variant 1</th><th>Variant 2</th><th>Full set</th><th>Rendered gate</th><th>Decision</th></tr></thead>
           <tbody>
             <tr>
-              <td><strong>Root · 1 → 2 → 3</strong><span>{liveCodexWorkerResult.productionBrowser.runs} Sol/medium run</span></td>
+              <td><strong>Root · delta v2</strong><span>{liveCodexWorkerResult.productionBrowser.runs} Sol/medium run</span></td>
               <td><strong>{displayMs(liveCodexWorkerResult.productionBrowser.firstVariantMs)}</strong><span>p95 {displayMs(liveCodexWorkerResult.productionBrowser.firstVariantP95Ms)}</span></td>
               <td><strong>{displayMs(liveCodexWorkerResult.productionBrowser.secondVariantMs)}</strong><span>independently acceptable</span></td>
               <td><strong>{displayMs(liveCodexWorkerResult.productionBrowser.allVariantsMs)}</strong><span>p95 {displayMs(liveCodexWorkerResult.productionBrowser.allVariantsP95Ms)}</span></td>
-              <td><strong>{liveCodexWorkerResult.renderedQuality.passed}/{liveCodexWorkerResult.renderedQuality.variants}</strong><span>production /bolder variants</span></td>
-              <td><span class="ks-tag" data-tone="ready">Default · replicate quality</span></td>
+              <td><strong>v2 passed</strong><span>{liveCodexWorkerResult.renderedQuality.passed}/{liveCodexWorkerResult.renderedQuality.variants} full-set variants</span></td>
+              <td><span class="ks-tag" data-tone="ready">Default · 54.5% faster v2</span></td>
+            </tr>
+            <tr>
+              <td><strong>Root · full source</strong><span>{liveCodexWorkerResult.fullSourceControl.runs} Sol/medium control</span></td>
+              <td><strong>{displayMs(liveCodexWorkerResult.fullSourceControl.firstVariantMs)}</strong><span>same full-card target</span></td>
+              <td><strong>{displayMs(liveCodexWorkerResult.fullSourceControl.secondVariantMs)}</strong><span>{displayMs(liveCodexWorkerResult.fullSourceControl.firstToSecondMs)} after v1</span></td>
+              <td><strong>{displayMs(liveCodexWorkerResult.fullSourceControl.allVariantsMs)}</strong><span>three full-source turns</span></td>
+              <td><strong>1/3</strong><span>production /bolder variants</span></td>
+              <td><span class="ks-tag is-neutral" data-tone="measure">Control · slower v2</span></td>
             </tr>
             <tr>
               <td><strong>Root · 1 → 3</strong><span>{liveCodexWorkerResult.twoCheckpointControl.runs} Sol/medium control</span></td>
@@ -442,7 +450,7 @@ const providerColumns = [
           </tbody>
         </table>
       </div>
-      <p class="live-performance-caption">Root-task samples prove target identity in the posted Generate payload. The earlier five-run heading sample remains useful only for transport latency. Next gates: replicate root-level rendered quality and reduce the {displayMs(liveCodexWorkerResult.productionBrowser.acceptToNextFirstVariantMs)} cold next-result wait; server pickup is already {displayMs(liveCodexWorkerResult.productionBrowser.serverPickupMs)}.</p>
+      <p class="live-performance-caption">Root-task samples prove target identity in the posted Generate payload. Speculative standby warmup was rejected: it slowed the next result to {displayMs(liveCodexWorkerResult.standbyExperiment.prewarmedNextFirstVariantMs)} versus {displayMs(liveCodexWorkerResult.standbyExperiment.controlNextFirstVariantMs)}. Next gates: replicate root-level rendered quality and reduce the cold post-Accept generation wait; server pickup is already {displayMs(liveCodexWorkerResult.productionBrowser.serverPickupMs)}.</p>
     </section>
 
     <section class="live-performance-section live-experiments-section ks-section" aria-labelledby="experiments-title" data-lab-panel="harness">
@@ -506,6 +514,7 @@ const providerColumns = [
               <dl>
                 <div><dt>First review</dt><dd>{displayMs(liveCodexWorkerResult.productionBrowser.firstVariantMs)}</dd></div>
                 <div><dt>Second review</dt><dd>{displayMs(liveCodexWorkerResult.productionBrowser.secondVariantMs)}</dd></div>
+                <div><dt>v1 → v2</dt><dd>{displayMs(liveCodexWorkerResult.productionBrowser.firstToSecondMs)}</dd></div>
                 <div><dt>All variants</dt><dd>{displayMs(liveCodexWorkerResult.productionBrowser.allVariantsMs)}</dd></div>
                 <div><dt>Accept → page ready</dt><dd>{displayMs(liveCodexWorkerResult.productionBrowser.acceptToResetMs)}</dd></div>
                 <div><dt>Accept → next Go</dt><dd>{displayMs(liveCodexWorkerResult.productionBrowser.acceptToNextGoDispatchMs)}</dd></div>
@@ -513,7 +522,7 @@ const providerColumns = [
                 <div><dt>Worker ready</dt><dd>{displayMs(liveCodexWorkerResult.coldWorkerReadyMs)}</dd></div>
                 <div><dt>Rendered gate</dt><dd>{liveCodexWorkerResult.renderedQuality.passed}/{liveCodexWorkerResult.renderedQuality.variants}</dd></div>
               </dl>
-              <p>Variant 2 is now independently reviewable and acceptable. The next gate is cold post-Accept generation, followed by replicated root-level rendered quality.</p>
+              <p>Variant 2 is independently reviewable, acceptable, and 54.5% faster with fenced source deltas. Cold post-Accept generation and replicated full-set quality remain the next gates.</p>
             </section>
           </aside>
         </div>