Record annotated and warm-cache evidence

Add repeated production annotation timing, real app-server cache telemetry, the context-delta decision, and correct --judge=false handling.\n\nAI-assisted: OpenAI Codex.

scripts/benchmark-live-codex-architecture.mjs

@@ -34,7 +34,7 @@ const timeoutMs = positiveInteger(args.timeout, 300_000);
 const outputPath = args.output ? path.resolve(ROOT, String(args.output)) : null;
 const profileIds = csv(args.profiles || 'direct-sol,cold-app-server,warm-app-server');
 const taskIds = csv(args.tasks || 'editorial-bolder,operations-polish,operations-annotated');
-const judgeEnabled = args.judge !== false;
+const judgeEnabled = args.judge !== false && args.judge !== 'false';
 const loadedEnv = loadBenchmarkEnv({ repoRoot: ROOT, explicitPath: args.envFile && path.resolve(args.envFile) });
 const skillPath = path.join(ROOT, '.agents', 'skills', 'impeccable', 'SKILL.md');
 const referenceDir = path.join(ROOT, 'skill', 'reference');

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

@@ -44,6 +44,9 @@ export const liveCodexWorkerResult = {
     annotatedQualityRuns: 3,
     coldInputTokens: 41066,
     coldOutputTokens: 2148,
+    warmRepeatInputTokens: 46973,
+    warmRepeatCachedInputTokens: 38656,
+    warmRepeatCacheRate: 0.823,
   },
   productionBrowser: {
     runs: 1,
@@ -55,6 +58,14 @@ export const liveCodexWorkerResult = {
     sourcePreflightMs: 52,
     validationMs: 10,
   },
+  annotatedBrowser: {
+    runs: 2,
+    firstVariantMs: 14316.21,
+    allVariantsMs: 45424.24,
+    comments: 1,
+    strokes: 1,
+    screenshotAttached: true,
+  },
   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.',

site/data/live-harnesses.ts

@@ -97,10 +97,10 @@ export const liveExperiments = [
   {
     rank: 8,
     title: 'Trim the producer contract — selected',
-    evidence: 'Progressive compact passed 5/5 for Claude Sonnet 4.6 and GPT-5.5. It was OpenAI’s fastest and cheapest progressive path; Claude’s other candidates all missed reliability.',
+    evidence: 'Progressive compact passed 5/5 for Claude Sonnet 4.6 and GPT-5.5. A repeated warm Sol turn cached 38,656 of 46,973 input tokens (82.3%) and completed in 21.8 s versus 26.0 s cold.',
     move: 'Keep the stable generation contract cacheable and send only the picked-element context plus the current action and identity lock.',
-    expected: 'Reduces prefill and output work without weakening the design brief.',
-    confidence: 'Five runs per candidate',
+    expected: 'Use provider caching instead of risky context deltas; always send current project and brand evidence.',
+    confidence: 'Candidate matrix + token telemetry',
   },
   {
     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>
-          A real Pick → <code>/bolder</code> → Go run produced variant 1 in {displayMs(liveCodexWorkerResult.productionBrowser.firstVariantMs)} and all three in {displayMs(liveCodexWorkerResult.productionBrowser.allVariantsMs)}. Accept returned the page in {displayMs(liveCodexWorkerResult.productionBrowser.acceptToResetMs)}; the next Go reached the independent worker {displayMs(liveCodexWorkerResult.productionBrowser.acceptToNextGoDispatchMs)} after Accept. In the matched full-task benchmark, warm app-server reached usable output in {displayMs(liveCodexWorkerResult.architectureComparison.warmAppServerFirstUsableMs)} versus {displayMs(liveCodexWorkerResult.architectureComparison.directFirstUsableMs)} direct; both passed 4/4. Production browser timing is one run.
+          A real Pick → <code>/bolder</code> → Go run produced variant 1 in {displayMs(liveCodexWorkerResult.productionBrowser.firstVariantMs)} and all three in {displayMs(liveCodexWorkerResult.productionBrowser.allVariantsMs)}. Accept returned the page in {displayMs(liveCodexWorkerResult.productionBrowser.acceptToResetMs)}; the next Go reached the independent worker {displayMs(liveCodexWorkerResult.productionBrowser.acceptToNextGoDispatchMs)} after Accept. The annotated browser path preserved its screenshot, comment, and stroke, then produced variant 1 in {displayMs(liveCodexWorkerResult.annotatedBrowser.firstVariantMs)} median. 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. Plain browser figures are one run; annotated figures are two.
         </p>
       </div>
     </section>