{"schema":"cain.e91.team.prereg.v2","registered_at":"2026-11-18T04:42:39Z","agreed_with":"session ec (message 2026-11-18)",
 "why":"team run 1 (bug_fix, remediation) was ceiling-saturated for Haiku (all configs 100%); families for run 2 chosen from DEV Haiku arm-D rates only (fault_localization 3/10, allocation 6/10), not from held-out",
 "model":"claude-haiku-5-5 via proxy :18891 (owner-approved scope: team experiment; same $5 cap)",
 "tasks":"held-out software.fault_localization (10) + business.allocation (10)",
 "configs":["single","three_identical","three_diverse","specialists","specialist_verifier","hierarchical","governed_swarm"],
 "budget_rule":"shared per-task tool budget: all agents and the verifier of one config draw on ONE probe/sample budget (fair vs single agent)",
 "verifier_evidence":{"fault_localization":"run_prefix(s-1)=True and run_prefix(s)=False (costs up to 2 probes from the shared budget)",
                      "allocation":"evaluate_plan proves FEASIBILITY only, not optimality; the verifier cannot catch suboptimal feasible plans"},
 "acceptance":"verticals.check only",
 "hypotheses":{"T4":"specialist_verifier or governed_swarm accepts more than single on fault_localization","T5":"no config beats single on allocation by more than 1 task (verifier sees feasibility only)","T6":"verifier false acceptance on allocation > 0 is possible and will be reported"},
 "no_changes_after_results":true}
