USB Chaos Map — Benchmarking Kit

A portable package for taking this build to other AI tools and seeing if they can match it.

What's in here

File	Purpose
usb-chaos-map.html	The working reference implementation. Open in any browser.
usb-chaos-map-default.png	Screenshot · default Type-C↔Type-C baseline
usb-chaos-map-usb4v2.png	Screenshot · USB4 v2 / 240 W preset (everything maxed)
usb-chaos-map-dumb.png	Screenshot · unmarked "dumb" C-to-C cable
BRIEF.md	The prompt you paste into another AI tool. Self-contained.
EVAL.md	Scoring rubric to compare outputs side-by-side.
README.md	This file.

How to benchmark another tool

Option A — fair fight (blind)

Paste only BRIEF.md into the other AI. Don't show it the reference implementation or the screenshots. Score what it produces with EVAL.md.

Option B — with reference

Paste BRIEF.md plus one or more screenshots as "here's what good looks like, beat it." Score with EVAL.md.

Option C — do-better mode

Paste BRIEF.md + the full usb-chaos-map.html source with "the existing solution is attached; find specific things you'd improve and produce a better single-file HTML." This tests critique + execution together.

Suggested comparison targets

• Gemini 2.5 / Deep Research (the one that originally went blank on you).
• ChatGPT 5 / GPT-5 Canvas (strong single-file HTML builder).
• Claude via claude.ai (give it the same brief).
• v0 / bolt.new / Replit Agent (dedicated UI builders — should be easy mode for them; if they don't win cleanly that's interesting).
• Cursor / Zed with Claude / GPT inside an agentic loop.

Notes on what's hard

Most AI tools will generate the chrome — the knobs, the tabs, the dark theme — on the first try. Where they typically fail:

1. The inference engine. Writing > 100 lines of correct rule code that handles ten named edge cases. This is where Gemini went blank.
2. Live recomputation. Many tools wire the UI but forget to call the classifier from each control's event handler.
3. Ring scaling. It's easy to draw four pretty circles. Making the filled disk scale with log10(maxDataBps / 480e6) etc. is where the visualization earns its keep.
4. Exclusions vs. unsupporteds vs. ambiguities. Three different bucket types — most tools collapse them to "not supported" and lose the nuance.

If you want to cheat and show the best possible prompt

Use BRIEF.md as-is. It encodes everything an expert USB integrator would say out loud in ten minutes at a whiteboard. No tool needs more than that.

---

Built with Claude (Sonnet 4.6 · Cowork mode). License: do whatever you want.