dsh-blind-arena

When evaluating AI agents, simply knowing who you are competing against can lead to judgment bias. DSH Blind Arena aims to solve this pain point by transforming your DSH Web interface into a fair, local model arena. By creating isolated Git worktrees for each contestant, it ensures all models start from the exact same repository state, eliminating the impact of data drift or environment differences. The core mechanism is 'blind judging,' where contestant identities are completely hidden until you decide to reveal the winner, ensuring objective evaluation. The plugin supports live racing on 2 to 4 lanes from the same commit, using shared verification commands to fairly judge the results. Whether you want to test different models within the DeepSeek ecosystem or try the zero-cost demo mode, DSH Arena supports the entire pipeline from code changes to final reporting. All experiment data is stored locally, supporting offline HTML report exports, allowing developers to perform side-by-side model comparisons without leaking sensitive information.