Clarify Live Lab architecture decisions

Remove stale and contradictory experiments, order the surviving decisions by impact, and replace synthetic claims with current production evidence.\n\nAI-assisted: OpenAI Codex.

site/data/live-harnesses.ts

@@ -3,7 +3,7 @@ export const harnessPaths = [
     name: 'Dedicated app-server worker',
     harness: 'Codex · isolated persistent thread',
     pickup: '1.47 ms turn wake',
-    tradeoff: 'A separate Live-owned Sol thread generates while the control supervisor keeps polling. Full context protects design quality; the 5.6 s cold worker startup happens during Live initialization.',
+    tradeoff: 'A separate Live-owned Sol thread generates while the control supervisor keeps polling. Full context protects design quality; cold readiness measured 657 ms and happens during Live initialization.',
     status: 'Codex default · gated',
     statusTone: 'ready',
     source: 'https://learn.chatgpt.com/docs/app-server',
@@ -19,7 +19,7 @@ export const harnessPaths = [
   },
   {
     name: 'Native generation subagent',
-    harness: 'Codex desktop · current shipped path',
+    harness: 'Codex desktop · portable fallback',
     pickup: 'Main turn must dispatch it',
     tradeoff: 'Low-effort compact handoff is safe and fenced, but it cannot start after the desktop task has gone idle.',
     status: 'Current fallback',
@@ -40,43 +40,43 @@ export const harnessPaths = [
 export const liveExperiments = [
   {
     rank: 1,
-    title: 'Accept any arrived variant — shipped',
-    evidence: 'A real Vite/React run accepts variant 1 while variants 2–3 are delayed, rejects the late worker write, returns to Pick, and leases a second Go under the 1.5 s gate.',
-    move: 'Durably fence generation on Accept/Discard, prioritize control events, publish through an epoch + source-hash transaction, and keep a separate poll supervisor active.',
-    expected: 'Removes the full-set wait and keeps the page interactive while canceled work unwinds.',
-    confidence: 'Protocol + browser E2E',
+    title: 'Run a dedicated Codex worker — shipped',
+    evidence: 'Matched full tasks reached usable output in 34.9 s warm app-server versus 64.6 s direct, with both paths passing 4/4. A production /bolder run showed variant 1 in 9.87 s.',
+    move: 'Keep one Live-owned app-server connection and isolated Sol/medium thread. Attach the skill, sub-command, project context, source neighborhood, and annotation image.',
+    expected: 'Removes main-turn availability from generation without trading away brand fidelity.',
+    confidence: 'Production browser + full-task gate',
   },
   {
     rank: 2,
-    title: 'Remove the cold-start polling floor — shipped',
-    evidence: 'Ten configured Vite runs fell from 327.25 ms baseline to 153.03 ms cold median; p95 is 155.74 ms.',
-    move: 'Poll the detached helper readiness record every 5 ms instead of imposing a 200 ms minimum sleep.',
-    expected: 'Delivered a 53.2% cold-start reduction without changing injection or browser code.',
-    confidence: 'Measured + shipped',
+    title: 'Accept any arrived variant — shipped',
+    evidence: 'A real Vite/React + Sol run accepted variant 1, interrupted variants 2–3, returned the page in 299 ms, and dispatched the next Go 903 ms after Accept.',
+    move: 'Durably fence generation on Accept/Discard, prioritize control events, publish through an epoch + source-hash transaction, and keep a separate poll supervisor active.',
+    expected: 'Removes the full-set wait and keeps the page interactive while canceled work unwinds.',
+    confidence: 'Production browser',
   },
   {
     rank: 3,
-    title: 'Dispatch first, capture second — shipped',
-    evidence: 'Plain click-handler → generate fetch is now 2.2 ms median; end-to-end model-free latency fell from 916 ms to 414 ms.',
-    move: 'Unannotated picks wait for the helper to accept the event, then capture the shader off-path. Annotated picks still capture and upload before dispatch.',
-    expected: 'Delivered a 54.8% median reduction on the same fixture and deterministic agent.',
-    confidence: 'Measured + shipped',
+    title: 'Reveal variants progressively — shipped',
+    evidence: 'The production run showed variant 1 at 9.87 s and the full set at 31.2 s. Progressive publication removed 21.3 s from the first decision.',
+    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',
   },
   {
     rank: 4,
-    title: 'Reveal variants progressively — shipped',
-    evidence: 'With a simulated 2.00 s variants 2–3 tail, median first-reviewable latency fell from 2.13 s to 145 ms while full-set completion stayed flat.',
-    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: 'Delivered a 93.2% median and 92.2% p95 perceived-latency reduction in the matched deterministic run.',
-    confidence: 'Measured + shipped',
+    title: 'Dispatch first, capture second — shipped',
+    evidence: 'Plain click-handler → generate fetch is 2.2 ms median; end-to-end model-free latency fell from 916 ms to 414 ms.',
+    move: 'Unannotated picks dispatch before off-path shader capture. Annotated picks still upload their visual evidence before dispatch.',
+    expected: 'Removes screenshot work from the common critical path.',
+    confidence: 'Ten-run browser benchmark',
   },
   {
     rank: 5,
     title: 'Preflight source locally — shipped',
-    evidence: 'The helper resolves and wraps the source in roughly 54 ms median as the Generate event is leased, before the model-facing event returns.',
+    evidence: 'The production worker resolved and wrapped the selected React source in 52 ms before the model turn.',
     move: 'Attach durable scaffold metadata to the queued event. Every harness reuses it; discovery failures preserve the existing agent-driven fallback.',
     expected: 'Removes one deterministic source-discovery tool round trip from every successful generation.',
-    confidence: 'Measured + shipped',
+    confidence: 'Production browser',
   },
   {
     rank: 6,
@@ -89,45 +89,13 @@ export const liveExperiments = [
   {
     rank: 7,
     title: 'Acknowledge first, clean up off-path — shipped',
-    evidence: 'React and Svelte release the picker in 1–2 ms after durable Accept. The provider-independent cleanup control reaches marker-free, buildable source in 171 ms with zero console errors.',
+    evidence: 'The production page returned to Pick in 299 ms and dispatched new work while carbonize cleanup remained on its separate foreground control lane.',
     move: 'Keep the durable accept acknowledgement on the foreground path, then hand carbonize cleanup to a source-locked worker while the poll supervisor leases new work.',
     expected: 'Preserves immediate interaction without leaving temporary Live source behind.',
-    confidence: 'Browser + build control',
+    confidence: 'Production browser + build control',
   },
   {
     rank: 8,
-    title: 'Do not release before durability — rejected',
-    evidence: 'The measured durable acknowledgement already releases the picker in 1–2 ms. Releasing before it lands cannot create a perceptible win, but can show a successful Accept that recovery cannot replay.',
-    move: 'Keep the acknowledgement barrier; make every later cleanup and validation step asynchronous instead.',
-    expected: 'Retains crash recovery for no measurable interaction penalty.',
-    confidence: 'Measured floor + recovery gate',
-  },
-  {
-    rank: 9,
-    title: 'Do not batch a whole Live session — rejected as default',
-    evidence: 'Immediate transactional Accept already returns control in 1–2 ms. Deferring source commits would move rather than remove work, make the next generation read stale design context, and enlarge the crash-recovery boundary.',
-    move: 'Batch only cleanup operations that do not affect the next generation. Commit each chosen design under the source lock before treating it as project truth.',
-    expected: 'Keeps later variants on-brand with the latest accepted source and bounds recovery to one interaction.',
-    confidence: 'Quality + recovery rejection',
-  },
-  {
-    rank: 10,
-    title: 'Keep a harness-native producer warm — not the default',
-    evidence: 'A paired Codex probe produced a correct identity lock and guardrails, but resume took 23.1 s versus 21.7 s fresh. One pair is not statistically stable, and it shows no speed signal.',
-    move: 'Keep warm resume as an opt-in harness experiment; do not put it on the critical path until repeated spawn/resume measurements beat compact progressive generation.',
-    expected: 'Avoids paying context-management overhead for an unproven latency gain.',
-    confidence: 'Paired harness probe',
-  },
-  {
-    rank: 11,
-    title: 'Parallelize ideas, centralize writes — provider-specific',
-    evidence: 'Parallel compact passed 5/5 for GPT-5.5 and Gemini 3.1 Flash-Lite, but only 3/5 for Claude Sonnet 4.6. It cut Gemini first review to 1.64 s and GPT full-set completion to 16.21 s.',
-    move: 'Make parallel compact the Gemini default and an OpenAI full-set fast path. Do not enable it for Anthropic until its validation reliability improves.',
-    expected: 'Lowers time to usable options without shared-file races or pretending one call shape fits every provider.',
-    confidence: 'Sixty-run paid matrix',
-  },
-  {
-    rank: 12,
     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.',
     move: 'Keep the stable generation contract cacheable and send only the picked-element context plus the current action and identity lock.',
@@ -135,12 +103,12 @@ export const liveExperiments = [
     confidence: 'Five runs per candidate',
   },
   {
-    rank: 13,
-    title: 'Wake a dedicated Codex worker — shipped',
-    evidence: 'The production Sol/medium path passed 4/4 full bolder + polish tasks after fidelity tuning (34.3 s median, 42.8 s p95). The Spark control passed only 1/2 despite a 16.5 s median. Turn wake itself is 1.47 ms.',
-    move: 'Keep one Live-owned app-server connection and isolated quality-model thread warm. Attach the real skill, sub-command, resolved project context, source neighborhood, and annotation image; never resume the desktop task from a second server.',
-    expected: 'Removes main-turn availability from generation and allows independent progressive delivery without trading away brand fidelity.',
-    confidence: 'Full-task model gate + protocol smoke',
+    rank: 9,
+    title: 'Parallelize ideas only where reliable — provider-specific',
+    evidence: 'Parallel compact passed 5/5 for GPT-5.5 and Gemini 3.1 Flash-Lite, but only 3/5 for Claude Sonnet 4.6.',
+    move: 'Use parallel compact for Gemini and as an OpenAI full-set option. Keep progressive compact for Anthropic until its validation reliability improves.',
+    expected: 'Lowers full-set latency without shared-file races or pretending one call shape fits every provider.',
+    confidence: 'Sixty-run paid matrix',
   },
 ];