A small, public, lighthearted research artifact: same prompt, same starting state, two Codex reasoning levels, two runnable Windows desktop widgets, and three blind LLM judges.
The short version: xHigh won all three blind judgments, averaging 44/50 versus 34/50 for Low. It also cost more time and tokens. That is the whole point of the experiment: compare the artifact, not just the vibe.
ARTICLE.md- a long-form X Article draft for publishing the story.PUBLISHING.md- notes for posting the article on X, including image upload order.TEASER_POST.md- a short X post to share the article.RESULTS.md- the measured results, judge scores, and findings.PROMPT.md- the exact prompt used for both runs.RUN_PROTOCOL.md- the execution and measurement protocol.RUBRIC.md- the judging rubric.MAPPING.md- the revealed A/B mapping after judging.apps/xhigh-system-pulse/- the xHigh .NET 8 WPF app.apps/low-system-pulse/- the Low PowerShell/WPF app.judges/- cleaned judge verdicts.assets/- postable charts, screenshots, and the article header image.data/- compact measurement data.
Raw transcripts, local orchestration files, and temporary runner logs are intentionally excluded from the public repo. They are useful locally, but they make a public research repo worse.
xHigh version:
cd apps\xhigh-system-pulse
dotnet run --project .\src\SystemPulse\SystemPulse.csproj -c ReleaseLow version:
cd apps\low-system-pulse
powershell -ExecutionPolicy Bypass -File .\SystemPulse.ps1Both target Windows 11. The xHigh version requires the .NET 8 SDK; the Low version uses built-in PowerShell and WPF.
| Metric | Low | xHigh |
|---|---|---|
| Average judge score | 34/50 | 44/50 |
| Wall-clock seconds | 499.5 | 793.9 |
| Output tokens | 10,839 | 27,032 |
| Stack | PowerShell + WPF | .NET 8 WPF |
| App CPU avg | 0.150% | 0.088% |
| App RAM | 158.1 MB | 128.9 MB |
This is not a benchmark suite. It is one controlled artifact comparison on one Windows widget task. Treat it as evidence, not gospel.