- Down {(improvement * 100).toFixed(1)}% from {displayMs(baselineFloor)} on the same fixture. Generate dispatch now starts in {displayMs(browserDispatch)} median.
+ Down {(improvement * 100).toFixed(1)}% from {displayMs(baselineFloor)} on the same fixture. Browser dispatch takes {displayMs(browserDispatch)} median.
Median values. Deterministic generation isolates Impeccable’s own floor.
+
Median product overhead with model time removed.
@@ -200,8 +173,8 @@ const providerColumns = [
- Result 01
-
Unannotated requests wait only for the helper to accept the event. Shader capture continues off-path; annotated requests still capture and upload first because the screenshot is semantic input.
+ Capture no longer blocks Go
+
Plain selections dispatch immediately. Annotated selections still upload the screenshot first because the model needs it.
Scroll horizontally to compare →
@@ -226,9 +199,9 @@ const providerColumns = [
Progressive delivery · Codex path
-
Review starts before generation ends
+
Variant 1 arrives first
-
Variant 1 becomes reviewable and acceptable immediately. Accept fences the old worker, releases the picker, and lets the poll supervisor lease the next Go while later variants unwind.
+
Review or accept the first result while later variants continue generating.
- Result 02
+ Measured improvement
Median first-reviewable latency fell {(progressiveGain * 100).toFixed(1)}%; p95 fell {(progressiveP95Gain * 100).toFixed(1)}%.
Full-set completion changed by {progressiveAllDelta.toFixed(1)} ms. This is a five-run model-backed comparison using {progressiveDeliveryResult.benchmark.provider} {progressiveDeliveryResult.benchmark.model} and a sanitized synthetic heading contract.
@@ -266,7 +239,7 @@ const providerColumns = [
Paid provider matrix · five runs per candidate
-
The right split depends on the provider
+
Choose delivery per provider
Sixty strict-gated runs. Brand, component, token, copy, source, and cleanup fidelity must all pass.
Ten isolated starts. App dependency installation and dev-server startup are excluded.
@@ -331,7 +304,7 @@ const providerColumns = [
- Result 03
+ Initialization result
Cold median fell {(initGain * 100).toFixed(1)}% to {displayMs(liveInitResult.cold.medianMs)}; p95 is {displayMs(liveInitResult.cold.p95Ms)}. Warm median is {displayMs(liveInitResult.warm.medianMs)}.
@@ -340,7 +313,7 @@ const providerColumns = [
Before / after / control
-
The common path is decisively faster
+
Plain selections skip capture
The annotated path remains a control: its screenshot must reach the agent before generation starts.
@@ -375,7 +348,7 @@ const providerColumns = [
Latency simulator
-
The gain survives slower models
+
Compare model and product time
Move model time to compare the old floor with the optimized product-side estimate.
@@ -405,7 +378,7 @@ const providerColumns = [
Harness delivery
-
Four ways to wake the work
+
Compare worker wake paths
Measure pickup before choosing architecture. Fast models cannot help while no model is running.
@@ -442,7 +415,7 @@ const providerColumns = [
Evidence-ranked decisions
-
What survived testing
+
Decisions and open bets
Shipped paths, rejected compromises, and unproven architectural bets are labeled separately.
@@ -464,7 +437,7 @@ const providerColumns = [
-
Measurement contract
+
How results are measured
bun run bench:live boots a real framework fixture and Chromium, drives Pick → Go → Cycle,
and records browser preparation, server pickup, scaffold, generation, write, and render boundaries.
@@ -474,9 +447,9 @@ const providerColumns = [