What it is
A small, honest side-by-side of a hosted decision model and an open one.
What it does: Asks one closed question about a headline and shows each engine's pick and confidence.
Why it belongs in the showcase: Clear, runnable demo of option-order sensitivity and cost per decision.
Caveats: API key and a large local download; toy task.
Evidence
Public MIT Python app (10★ at intake) created 2026-10-16. README documents uv quick start, TYPESAFE_API_KEY in .env, an `onion-or-not download` step for the official Qwen3-8B GGUF (Q4_K_M ~5 GB default), and an "ask with the options swapped" toggle to expose order flips. It states about 300 input tokens per question at $0.042 per million and shows per-answer cost in the UI.
Caveats
- Jev card needs your own TYPESAFE_API_KEY; local models need a ~5 GB download
- Satire detection is a demo task, not a fact-checking service
- Compares three engines; results depend on headline wording
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.