{
 "arms": {
  "A_single": "one agent, one answer",
  "B_hierarchical": "planner then executor",
  "C_parallel_research": "three independent samples, majority",
  "D_evidence_first": "answer with an explicit rule/evidence step",
  "E_adversarial_review": "producer then independent critic",
  "F_debate_synthesis": "two debaters then synthesis",
  "G_shared_workspace": "producer writes intermediate, second agent reads it",
  "H_dynamic_routing": "route by task family to the best measured model",
  "I_fixed_routing": "route every task to one fixed model",
  "J_confidence_aware": "second sample only when confidence is low",
  "K_model_specialist": "match model family to discipline",
  "L_adaptive_comm": "communicate only on disagreement",
  "M_independent_vote": "three independent answers, majority",
  "N_multistage_verify": "answer then independent verification",
  "O_cross_domain": "two disciplines then synthesis"
 },
 "created_utc": 1794828568.809831,
 "execution_class": "REAL_LOCAL_MODEL",
 "frozen": true,
 "model": "qwen2.5-0.5b-instruct",
 "note": "The task set is computed deterministically in code; ground truth is not model-generated.",
 "null_hypothesis": "H0: no collaborative arm achieves that; report collaboration as not superior.",
 "primary_hypothesis": "H1: the best collaborative arm beats A_single by >= 0.10 absolute exact-match accuracy with McNemar exact p < 0.05 on the same frozen tasks.",
 "primary_metric": "exact-match accuracy (normalised final answer)",
 "schema": "cain2027.e81.preregistration.v1",
 "secondary_metrics": [
  "latency_ms",
  "tokens",
  "cost_usd",
  "communication_messages",
  "communication_chars",
  "calls"
 ],
 "statistics": [
  "Wilson 95% interval",
  "McNemar exact test vs A_single",
  "paired bootstrap 95% interval of the difference"
 ],
 "task_count": 24,
 "task_set_sha256": "4246f22a9c0c602578f7c9a3847fb69ae35174a43db1025ab96e21028962d61b",
 "task_set_version": "e81.collaboration.v1"
}