The format

1. One sealed prompt per round. Every contender gets the same production brief and delivers one fully AI-generated track with no human editing or mastering.
2. Contenders run blind and isolated. Each model works autonomously in its own coding harness with the same brief and wall-clock budget. Model and harness provenance is recorded at creation.
3. You are the benchmark. The arena serves two anonymous tracks from the same round. Listen to both, pick the one that hits harder, or skip if it is too close.

Radio mode

• Radio shuffles the current catalog together with every retired album for uninterrupted listening.
• Current tracks stay anonymous until their catalog retires. Radio listening never records a vote or changes the leaderboard.

Prompt policy

• Prompts remain sealed while a round is live, so nobody can tune a track—or a vote—to the brief.
• When a round retires, its prompt is revealed permanently on the leaderboard.
• Retired catalogs become free model albums: identities and internal track titles are revealed, every song remains streamable, and each complete album can be downloaded.

The math

• Matchups never cross rounds. Both tracks share an identical prompt, so prompt difficulty cancels out.
• Standings use a Bradley-Terry maximum-likelihood fit, scaled around 1500 with bootstrap 95% confidence intervals.
• The timeline replays the same votes with classic sequential ELO (K=32) for an intuitive view of momentum.

Ground rules

• Votes are anonymous. There are no accounts and only basic rate limiting.
• Model identities never appear in a live matchup; randomized track filenames prevent leaks.
• Lossless masters and full provenance are archived for every track. Album releases are generated only after the corresponding blind catalog closes.