Eval arena
A short leaderboard of prompt and model candidates scored by three judges: run a new candidate and watch 48 cases fill while its row climbs the board and its dot glides across quality versus cost, then open each regression side by side.
Data and insightNo. 2382 variationsFree with an account
Playing on its own
Variation
- The idea
- Evals as the test suite for prompts: a run you can watch, ranked against cost, with every regression one click from a diff and the judge's reason
- Autoplay
- Run eval on Wren 3 cite: the case grid fills in a wave, the row climbs from last to second while its dot glides on the chart; a 3 regressions link appears; the diff sheet opens with both answers and the changed words marked; the cursor steps to the next regression, confirms it, closes the sheet and switches the judge to Faithful so the new candidate leads.
- When you click
- Run and Stop (R); a judge switch picks the score column and the order; hover a row or a dot to link them; set any candidate as the baseline; arrow through the case grid and press Enter on a red cell; in the diff sheet step with Left and Right, confirm or dismiss a regression, Esc closes and returns focus. Phone: board rows, the run strip and the chart stack, the diff is a full sheet.
- Look
- flat hairline cards, one score column, FLIP reorder, cell pop, sans with mono scores type
- Variations
- Graphite, Lab
- Files
- index.html, style.css, script.js and prompt.md. No framework, no CDN, works offline.
The prompt and the files
Paste the prompt into Claude, ChatGPT, Lovable or any coding agent to rebuild it in your own product, or take the three files as they are.
Loading the files