{
 "version":"1.0-before-execution",
 "question":"How do direct thread delivery and duplicate identifiers affect the routing work and correctness of the published Super-agent domain Engine?",
 "source":"deepseek-ai/dsh-superagent-prototype",
 "commit":"6337fc823ce20251c8de8085948165c81aaa701e",
 "scope":"Unmodified src/core with upstream MemoryRepositories. Assessor, Executor, Policy and Outbox are explicit deterministic adapters, not real models, tools, authorisation adapters or channels.",
 "dataset":{"threads":12,"initial_requests":12,"human_followups_per_thread":10,"unique_events":132,"order_seeds":[17,42,73],"extra_copies_per_event":[0,1,2],"drain_batch_sizes":[1,8]},
 "variants":["native thread subscriptions","subscriptions removed using Engine.unsubscribe; a deterministic oracle calls Engine.attach for every routed follow-up, then removes subscriptions again"],
 "runs":36,
 "metrics":["assessor request count after initial task creation","unique follow-ups delivered to their expected ticket","repeated deliveries of the same input","wrong-ticket deliveries","duplicate receive return count"],
 "boundary_scenarios":["same id dedup","same text new id control","human stop","bot after completion","human resume","revoked resume","two owners handing back","context-only subscription after completion","paused evaluation timeout","message text does not mutate mandate"],
 "no_claims":["LLM accuracy","token savings","latency speedup","durable persistence","sandbox security","DeepSeek original benchmark reproduction","CapyAgent comparative win"]
}