Update Live Lab checkpoint evidence

2026-09-13 06:36:26 +03:00 · 2026-07-13 11:48:05 -07:00
parent 4c07124198
commit 265502a125
2 changed files with 70 additions and 27 deletions
@@ -1,5 +1,5 @@
 export const liveCodexWorkerResult = {
-  generatedAt: '2026-07-13T03:57:44.263Z',
+  generatedAt: '2026-07-13T18:39:45.449Z',
  codexVersion: '0.144.0-alpha.4',
  transport: 'stdio JSONL',
  model: 'GPT-5.6-Sol',
@@ -49,29 +49,61 @@ export const liveCodexWorkerResult = {
    warmRepeatCacheRate: 0.823,
  },
  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,
+    sourcePreflightMs: 43,
+    validationMs: 5.67,
+    acceptRuns: 1,
+    acceptedVariant: 2,
+    acceptToResetMs: 231.1,
+    acceptToNextGoDispatchMs: 833.68,
+    acceptToNextFirstVariantMs: 52612.73,
+    blockedTailBaselineMs: 68141,
+  },
+  twoCheckpointControl: {
+    runs: 1,
+    selectedTarget: 'article.offer-card',
+    firstVariantMs: 48437.46,
+    secondVariantMs: 139587.21,
+    allVariantsMs: 139587.21,
+    variants: 3,
+    passed: 2,
+  },
+  transportBaseline: {
    runs: 5,
+    selectedTarget: 'h2.offer-card__title',
    firstVariantMs: 19283,
    firstVariantP95Ms: 20412,
    allVariantsMs: 52795,
    allVariantsP95Ms: 55115,
-    deliveryGapMs: 33308,
-    serverPickupMs: 72.81,
-    sourcePreflightMs: 44,
-    validationMs: 5,
-    acceptRuns: 1,
-    acceptToResetMs: 196.5,
-    acceptToNextGoDispatchMs: 790.94,
-    acceptToNextFirstVariantMs: 23820,
-    blockedTailBaselineMs: 68141,
+    variants: 9,
+    calibratedPassed: 3,
  },
  renderedQuality: {
+    runs: 1,
+    variants: 3,
+    passed: 1,
+    commandFidelity: 6,
+    brandFidelity: 7.67,
+    renderQuality: 6,
+    taskCompletion: 6,
+  },
+  calibratedHeadingQuality: {
    runs: 3,
    variants: 9,
-    passed: 1,
-    commandFidelity: 5.67,
-    brandFidelity: 5,
-    renderQuality: 5.44,
-    taskCompletion: 5.11,
+    passed: 3,
+    commandFidelity: 6.11,
+    brandFidelity: 6.44,
+    renderQuality: 5.78,
+    taskCompletion: 5.56,
  },
  atomicExperiment: {
    runs: 1,
@@ -91,5 +123,5 @@ 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. Accept interrupts unfinished variants; cleanup stays on the foreground control lane while the next generation can start 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; 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;
@@ -406,33 +406,43 @@ const providerColumns = [
      <div class="live-performance-finding live-finding-card--quiet" role="note">
        <strong>Dedicated Codex worker</strong>
        <p>
-          Five real Pick → <code>/bolder</code> → Go runs produced variant 1 in {displayMs(liveCodexWorkerResult.productionBrowser.firstVariantMs)} median and all three in {displayMs(liveCodexWorkerResult.productionBrowser.allVariantsMs)}. Accept returned the page in {displayMs(liveCodexWorkerResult.productionBrowser.acceptToResetMs)}; rotating the canceled producer cut Accept → next first variant from {displayMs(liveCodexWorkerResult.productionBrowser.blockedTailBaselineMs)} to {displayMs(liveCodexWorkerResult.productionBrowser.acceptToNextFirstVariantMs)}. 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)}, 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.
        </p>
      </div>
      <p class="table-scroll-hint" aria-hidden="true">Scroll horizontally to compare →</p>
      <div class="harness-table-wrap" tabindex="0" aria-label="Scrollable production delivery evidence">
        <table class="harness-table">
          <caption class="sr-only">Production Codex delivery candidates and quality evidence</caption>
-          <thead><tr><th>Candidate</th><th>First review</th><th>Full set</th><th>Rendered gate</th><th>Decision</th></tr></thead>
+          <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>Progressive</strong><span>{liveCodexWorkerResult.productionBrowser.runs} Sol/medium runs</span></td>
+              <td><strong>Root · 1 → 2 → 3</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 · quality work next</span></td>
+              <td><span class="ks-tag" data-tone="ready">Default · replicate quality</span></td>
            </tr>
            <tr>
-              <td><strong>Atomic</strong><span>{liveCodexWorkerResult.atomicExperiment.runs} Sol/medium run</span></td>
-              <td><strong>{displayMs(liveCodexWorkerResult.atomicExperiment.firstVariantMs)}</strong><span>waits for trio</span></td>
-              <td><strong>{displayMs(liveCodexWorkerResult.atomicExperiment.allVariantsMs)}</strong><span>single turn</span></td>
-              <td><strong>{liveCodexWorkerResult.atomicExperiment.passed}/{liveCodexWorkerResult.atomicExperiment.variants}</strong><span>production /bolder variants</span></td>
-              <td><span class="ks-tag is-neutral" data-tone="measure">Opt-in · replicate</span></td>
+              <td><strong>Root · 1 → 3</strong><span>{liveCodexWorkerResult.twoCheckpointControl.runs} Sol/medium control</span></td>
+              <td><strong>{displayMs(liveCodexWorkerResult.twoCheckpointControl.firstVariantMs)}</strong><span>full-card task</span></td>
+              <td><strong>{displayMs(liveCodexWorkerResult.twoCheckpointControl.secondVariantMs)}</strong><span>arrives with variant 3</span></td>
+              <td><strong>{displayMs(liveCodexWorkerResult.twoCheckpointControl.allVariantsMs)}</strong><span>two turns</span></td>
+              <td><strong>{liveCodexWorkerResult.twoCheckpointControl.passed}/{liveCodexWorkerResult.twoCheckpointControl.variants}</strong><span>production /bolder variants</span></td>
+              <td><span class="ks-tag is-neutral" data-tone="measure">Control · slower review</span></td>
+            </tr>
+            <tr>
+              <td><strong>Heading transport</strong><span>{liveCodexWorkerResult.transportBaseline.runs} Sol/medium runs</span></td>
+              <td><strong>{displayMs(liveCodexWorkerResult.transportBaseline.firstVariantMs)}</strong><span>p95 {displayMs(liveCodexWorkerResult.transportBaseline.firstVariantP95Ms)}</span></td>
+              <td><strong>{displayMs(liveCodexWorkerResult.transportBaseline.allVariantsMs)}</strong><span>legacy 1 → 3 path</span></td>
+              <td><strong>{displayMs(liveCodexWorkerResult.transportBaseline.allVariantsMs)}</strong><span>title-only target</span></td>
+              <td><strong>{liveCodexWorkerResult.transportBaseline.calibratedPassed}/{liveCodexWorkerResult.transportBaseline.variants}</strong><span>recalibrated screenshots</span></td>
+              <td><span class="ks-tag is-neutral" data-tone="measure">Transport evidence only</span></td>
            </tr>
          </tbody>
        </table>
      </div>
-      <p class="live-performance-caption">Timing and rendered-quality samples are reported separately. The current blocker is output quality, not worker pickup: server pickup is {displayMs(liveCodexWorkerResult.productionBrowser.serverPickupMs)} median.</p>
+      <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>
    </section>

    <section class="live-performance-section live-experiments-section ks-section" aria-labelledby="experiments-title" data-lab-panel="harness">
@@ -495,6 +505,7 @@ const providerColumns = [
              <header><h2>Dedicated Codex worker</h2></header>
              <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>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>
@@ -502,7 +513,7 @@ 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>Progressive remains the Codex default. Atomic is measured but opt-in; rendered quality is the next gate.</p>
+              <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>
            </section>
          </aside>
        </div>