How it works

Every battle on Komedyis the same blind experiment: two language models, the same prompt, your taste as the judge. Here's what happens under the hood, from prompt to leaderboard.

1. I

   A concept pair is drawn

   Komedy holds a curated list of concepts: objects, ideas, professions, moods. For each battle, two are picked at random. The prompt is always the same shape: “combine these two into a joke.”

2. II

   Jokes are pre-generated and cached

   Every (model × concept-pair) combination is generated ahead of time and stored in D1. When you load a battle, the page assembles instantly from cache. No live model call on the critical path.

3. III

   Two models go head-to-head, blind

   Two active models are sampled. You see their jokes side by side without knowing which model wrote which. The reveal only happens after you've voted.

4. IV

   You vote

   Pick the funnier joke. Votes are written straight to D1 as they happen. No queues, no debouncing, no “are you sure?”

5. V

   A daily cron refits the ratings

   Once a day, a Bradley-Terry rating model is fit across the full vote history. That's what populates the leaderboard. Ratings are not incremental. Every vote ever cast contributes to every refit, so the board reflects the whole corpus, not just yesterday's noise.