changer-changer/dsh-blind-arena

README 摘要

DSH Arena Blind, fair, local Agent arena inside DSH Web. Same task. Same commit. Isolated worktrees. Shared verification. Judge before you reveal. DSH Arena turns your DeepSeek Harness (DSH) Web into a model-vs-model fight club: pick a task, pick 2–4 contestants (any provider/model DSH supports), and let them race on the exact same repository state. Lane identities are kept secret until you decide — no label bias, no prompt leaks, no cherry-picked baselines. Why this exists Benchmarks lie. Leaderboards are curated. When you evaluate an agent, the simple act of knowing which model you're looking at biases everything downstream: your prompts, your patience, your verdict. DSH Arena removes that bias structurally: - One repository state — every lane starts from the same commit, in its own git worktree. No drift, no stale forks. - Random lane assignment — participants are shuffled onto lanes A/B/C/D at creation. The host knows; the wire doesn't. - Identity-free protocol — the public experiment record contains no participant identity. Identities live in a server-side secrets.json and merge into the view only after you call reveal . - Shared verification — the same user-defined verify com…

分类