SoulSim replay and a separate29-agent private-world account Reviewed September6,2026 UTC. Public read-only research. NEW CHINESE-COMMUNITY PROJECT https://www.v2ex.com/t/1227531 July15 announcement by xiaoxuz links https://github.com/xiaoxuz/SoulSim and https://xiaoxuz.github.io/SoulSim/ . The author describes a world-simulation workspace with persistent characters, interventions and follow-up conversations. This supplies Chinese-language operator context, not Chinese lab attribution. Pinned source revision:68a8a5111a030e5ce205e2b73ee5db537437bea4. GitHub recursive tree returned complete. Four source files were fetched, not executed. WHAT THE PUBLIC DEMO ACTUALLY DOES README explicitly describes a browser replay requiring no backend, database or LLM service. Source frontend/lib/demoApi.ts confirms this: it imports fixed demoData, constructs streaming events with delays, and returns local cloned records. Its groupEvents function takes the first three demo agents and formats a generic response around the user's first20characters and the role type. Thus apparent fresh group-chat responses in this mode are templates, not evidence of live agents. frontend/lib/demoData.ts contains five characters and a three-step run with five response entries per step,15total. World ID2d25b75c-2a87-4c42-a7c2-80e20dc4e165; run ID58ffee8a-6bc2-450e-9331-e1604a1c418c. The world concerns adapting to AI-driven job displacement. Embedded started/finished fields areJuly14,2026 17:34:33.752932 and18:13:42.304058 atUTC+08. These are publisher-supplied record fields, not an independent39-minute observation. This pass does not establish whether the fixed scenario was originally model-generated, edited or hand-authored. Both the announced demo URL and three inspected V2EX threads returnedHTTP200 to direct requests. Web-reader cache failures for twoURLs were not an actual direct-access block. No demo buttons were activated. REAL BACKEND CAPABILITY, SEPARATE FROM REPLAY backend/app/chat/group_agent.py contains a synchronous event queue. A human message is saved, candidate agents selected, relevance assessed, and responses generated through the LLM-client wrapper. Saved assistant responses are appended as new pending events, allowing another agent to respond. The sender is excluded from its own event's candidate list. Recent history and a compact discussion state are passed forward. Repetitive replies can be retried once and then skipped. The checked function bounds processed events using max(3,max_rounds)*max(1,number_of_agents), stops after three quiet events, and checks whether a newer human message has arrived. This is a processed-event budget, not an exact cap on model calls or output messages; each event can select multiple candidates and perform extra relevance/state/retry calls. No concurrency or deployment behavior was tested. engine/loop.py was also preserved, but not fully audited in this pass. These are implemented mechanisms for bounded character interaction. They do not demonstrate an unattended public swarm. SEPARATE29-AGENT OPERATOR ACCOUNT https://www.v2ex.com/t/1201182 comment15, March26: dimlau describes29 agents with memories and interpersonal interactions, observed through simulated letters and a simulated social feed. They say the puzzle-game direction was largely abandoned because characters pursued their own goals. https://www.v2ex.com/t/1208164 comment8, April24: the same account again describes29 NPCs, letter-based interaction and relatively stable remembered relationships. Comment12, May3 explicitly says it is running privately for personal observation and has no public operation. This is repeated first-person testimony over time, not multiple independent witnesses. No run log or project repository was linked in these inspected comments. The private-operation disclosure explains why absence of a public stream is unsurprising. Do not infer a hidden public endpoint or pursue private access. It is unrelated to SoulSim unless new evidence links them; none currently does. ASSESSMENT AND NEXT SEARCH Both leads broaden the search beyond coding teams: simulated societies and persistent fictional characters are another place Chinese operators describe agent-to-agent activity. Public replay data, actual backend capability and private-run testimony must remain separate evidence classes. Neither identifies an escaped swarm or XZ. Pending: chenhg5/agencycli, announced at V2EX1206945; search for intentionally published workflow outputs and issue discussions. FastClaw announcement V2EX1222063 is another unvalidated infrastructure lead. No accounts, messages, model calls or target executions were created during this round. INFRASTRUCTURE The threeV2EX pages and SoulSim demo were reachable from the present server. Additional mainland access is useful for the separately documented blocked Chinese pages, but will not turn replay data into live evidence or reveal a privately run world. An existing always-on mainland broadband computer with a dedicated browser profile and remote access is the most useful comparison environment. Rough engineering starting point:2CPU cores,4GBRAM,40-70GBdisk; noGPU needed. These are estimates, not a vendor quote or guarantee of access. Public-page exports with source URL and capture time can also help without provisioning hardware. See website access.html for exact blocked examples and failure distinctions. PRESERVATION 333-private contains full tree, four pinned source files, demo aggregate, three publicV2EX HTML/text captures, hosted-demo HTML/text, access results, publication verification and checksums. No private credentials or target configuration files were fetched.