{
 "B_single_on_same_tasks": {
  "accuracy": 0.375,
  "correct": 6,
  "tasks": 16
 },
 "T2_tool_first": {
  "accuracy": 0.25,
  "correct": 4,
  "domains": [
   "mathematics",
   "data_analysis",
   "scientific_research",
   "planning"
  ],
  "raw_rows": "artifacts/e79/bench_iterations/T2_tool_first.jsonl",
  "tasks": 16,
  "tokens": 1774,
  "tool_denied": 5,
  "tools_used": 15
 },
 "delta_vs_single": -0.125,
 "execution_class": "LOCAL_MODEL_INFERENCE + DETERMINISTIC_TOOL_EXECUTION",
 "grants_authority": false,
 "interpretation_note": "T2 removes the second instruction. If T2 beats B on the same tasks, the weak T arm was a protocol failure. If it does not, the tool did not help this model on these tasks.",
 "model_scale_same_tasks": {
  "accuracy": 0.375,
  "correct": 6,
  "execution_class": "LOCAL_MODEL_INFERENCE",
  "model": "Qwen/Qwen2.5-1.5B-Instruct",
  "raw_rows": "artifacts/e79/bench_iterations/B_qwen1_5b.jsonl",
  "tasks": 16,
  "tokens": 1710
 },
 "schema": "cain2027.e79.protocol_iterations.v1"
}