{"version":"configuration-source-6","evidencePolicy":"documented","snapshotId":"sha256-3848b85526102e7a985d53044e8baa86203d9d16f2affbb4a93bcb265f95b8fb","effortDatasetId":"sha256-49305f3d640f7387a764d95d62f10bbce44aeff4a4012ddba973eddf1e57a707+sha256-443256461e035ccdee585c01fdbd68b4b26545d45e2bd6b2bd93e1e913a2d41a+sha256-32c6374f8a0cdb9cd64f17244f1a697dd35d31cff071002599f3e1226d1e027d+sha256-9b7bfe52f0cba33df0362a330cb12551ea2f5853bf843931b08dad98e5d8d15c+sha256-c4f243913106bee769c57849d3cb4662fee3097206d30d3c8d4d37d687f8e556+sha256-c6f9cb8b67b4289e79e621b65d529e5f57eee55d89d86915f397859092051dad+sha256-78cb5ce46bffe96a698a9a7117551eafb90b2c3d05066722f4a703be2680aa55+sha256-e34a2de55476481fa69ea734e4027620fbc14df27f896c3bcec6c07275b0da42","scope":"All collected observations, not a claim that every public result has been collected. Contributing means positive graph weight, not sufficient evidence for a rank.","models":[{"id":"hy4-preview","collected":7,"contributing":0,"admittedWithoutWeight":0,"excluded":7}],"observations":[{"id":"b904deb8ba6d1f0fbb5abf52","modelId":"hy4-preview","configurationId":"hy4-preview--effort-high","sourceUrl":"https://surgehq.ai/benchmarks/dayjob-healthcare","status":"excluded","reason":"Evaluator has not been reviewed for independent admission"},{"id":"319cd5622cf0976b6dbcc4b7","modelId":"hy4-preview","configurationId":"hy4-preview--effort-high","sourceUrl":"https://surgehq.ai/benchmarks/dayjob-finance","status":"excluded","reason":"Evaluator has not been reviewed for independent admission"},{"id":"65a34d1fa2ecb6776ad932e8","modelId":"hy4-preview","configurationId":"hy4-preview--effort-high","sourceUrl":"https://surgehq.ai/benchmarks/sudo-l7","status":"excluded","reason":"Evaluator has not been reviewed for independent admission"},{"id":"279ad2ba50cd782c323458b5","modelId":"hy4-preview","configurationId":"hy4-preview--effort-unspecified","sourceUrl":"https://arena.ai/leaderboard/agent","status":"excluded","reason":"Effort is not identified"},{"id":"s-dayjob-dayjob-finance-hy4-preview-high-d1a93089bde9","modelId":"hy4-preview","configurationId":null,"sourceUrl":"https://surgehq.ai/benchmarks/dayjob-finance","status":"excluded","reason":"Supporting evidence outside the reviewed capability core"},{"id":"s-dayjob-dayjob-healthcare-hy4-preview-high-d1a93089bde9","modelId":"hy4-preview","configurationId":null,"sourceUrl":"https://surgehq.ai/benchmarks/dayjob-healthcare","status":"excluded","reason":"Supporting evidence outside the reviewed capability core"},{"id":"s-lmarena-agent-hy4-preview-unspecified-2ac93e0d4bf1","modelId":"hy4-preview","configurationId":null,"sourceUrl":"https://arena.ai/leaderboard/agent","status":"excluded","reason":"Supporting evidence outside the reviewed capability core"}]}