Add a separate DSH arm while preserving all baseline and DeepAgents archives.
The fixed 1000-task selection includes 87 preselected technical supplements; all 1087 attempts and unknown usage are retained. The API alias is deepseek-v4-pro; served weights and strict official protocol parity are unverified. Full replication: 37.2%. HF-style metrics: 69.9%, 55.6%, 68.3%, 59.8%.
Publication reads sealed scores without model calls or rescoring. The display preserves existing paired comparisons and links to a pinned archive commit.
Refresh leaderboard with single-pass Agent and DeepAgents comparisons (#4)
