File size: 1,817 Bytes
63eb615 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 | {
"protocol_version": 1,
"status": "frozen-before-first-run",
"frozen_router_commit": "06b5381b1521c3db34944ac9ab2b8cb07be438f8",
"frozen_scope": "agentweave/",
"external_benchmark": {
"name": "General-AgentBench",
"repository": "cxcscmu/General-AgentBench",
"commit": "35f5c027c31ddcb3366b28674c6cb2957460c0e2"
},
"task_source": "published baseline trace user prompts only",
"canonical_trace_directories": {
"mathhay": "mathhay_Claude-Haiku-4.5_distraction_all",
"mcpbench": "mcpbench_Claude-Haiku-4.5_distraction_all",
"search": "search_Claude-Haiku-4.5_distraction_all",
"swebench": "swebench_Claude-Haiku-4.5_distraction_100",
"tau2bench": "tau2bench_Claude-Haiku-4.5_distraction_all",
"terminalbench": "terminalbench_Claude-Haiku-4.5_distraction_100"
},
"label_visibility": "benchmark labels are withheld until after requirement analysis and family prediction",
"prediction_rule_order": [
"mcpbench: mcp or tool-use capability",
"terminalbench: operating-system or shell capability",
"tau2bench: compliance capability/domain",
"swebench: coding, backend, frontend, database, or sql capability",
"search: research, retrieval, or knowledge-graph capability",
"mathhay: otherwise"
],
"primary_metric": "family Hit@1",
"secondary_metrics": ["macro family accuracy", "per-family accuracy", "mean analyzer confidence", "confusion matrix"],
"anti_tuning_rule": "The benchmark fails if any file under agentweave/ differs from frozen_router_commit. Results must be published without changing the frozen router for this evaluation.",
"evidence_boundary": "This is an untouched external family-routing generalization test over published General-AgentBench prompts. It is not the native General-AgentBench end-to-end task score."
}
|