Receipt-driven benchmark for how coding models handle underspecified single-HTML browser tasks.
The live suite is Rollercoaster with a frozen A/B/C prompt ladder: three cells per model. A is the raw task. B adds 20 words of autonomy. C adds 20 words of showcase pressure. The model still chooses architecture, interaction, look, and scope. Ant Colony and Fireworks are suspended; their prompts and published takes are preserved in .archives. More benches can be restored or added later without changing that ladder.
Live gallery: benchmark.gvaste.dev.
Requires Vite+ (vp) and pnpm@12.0.0.
vp install
vp check
vp test
vp run bench:doctor
vp run devvp run dev serves the gallery locally.
vp run bench -- list
vp run bench -- plan --models grok-4.6 --benchmarks rollercoaster --levels A,B,C --attempts 2 --adapter agent --harness cursor
vp run bench -- gallery --run <run-id>CLI, run modes, and gallery publish: docs/commands.md.
MIT.