Pelican Arena: Blind-Testing Models on a Bicycle-Drawing Challenge

Field notes

The benchmark uses a deliberately awkward SVG assignment—drawing a pelican riding a bicycle—to test whether models can construct spatially coherent vector code rather than reproduce an obvious existing snippet. The visible interface pairs two concealed submissions, supports keyboard voting and ties, and says identities appear after the vote. It also names a dynamic Elo ladder, an eight-model collection, a gallery, and API checks for latency, throughput, and alleged chain-of-thought impersonation.