docs(exploration): prior-art L2/L3 — Chinese-community supplement

[cyber-ayi]

Summary

Supplements exploration/prior-art-l2-l3.md with five concrete items the original 2026-06-03 survey missed, all surfaced by a Chinese-community-focused research pass.

Most load-bearing finding: GenerativeAgentsCN (x-glacier, MIT, 463⭐) — Smallville zh reimplementation that empirically validates 25 LLM agents running on local Ollama + Qwen3-4B / DeepSeek-R1. Materially de-risks the local-cost question if/when a multi-agent substrate is considered. Companion exploration note (substrate-evennia-multi-agent) is being opened in parallel and links here.

What's added

L2 table (one row) — AgentScope (Alibaba DAMO, Apache-2.0). msghub multi-agent broadcast + 7-agent Werewolf template; closest off-the-shelf "drop N agents into a room" runtime. Flagged for spike vs ElizaOS-core only if the multi-agent substrate option is pursued.

L3 table (three rows):

• GenerativeAgentsCN — concrete N=25 local-LLM cost evidence.
• AgentVerse simulation track (OpenBMB/Tsinghua, arxiv 2308.10848) — third multi-agent-emergence reference besides Smallville/PIANO.
• EconAgent (Tsinghua, ACL'24 Outstanding) — 100 LLM × 20-yr macro sim reproducing stylized facts. Strongest evidence that long-horizon multi-LLM sims can stay coherent (independent of the objective).

New "Chinese-community supplement" subsection — items that don't fit the L2/L3 dichotomy:

• CharacterGLM-6B (THU CoAI + Lingxin, EMNLP'24) — NPC local-model candidate.
• Chinese RP corpora (ChatHaruhi 54k / CharacterEval 1785·77 / RoleBench / SuperCLUE-Role) — reusable as persona-fidelity / believability evaluators + RAG corpus.
• Wuxia-MUD lib assets (pkuxkx.net wiki + mudcore) — 30 yrs of LPMud zh content as RAG corpus if a zh-setting substrate is chosen (setting choice deferred).
• AI-companion product observation — 筑梦岛 / 猫箱 / 星野 / Tavo all ship "multi-AI in one shared scene"; zh-market consumer validation of multi-agent co-presence UX (closed-source; informative only).
• One MUD × LLM lead — mud.ren/threads/436 "Yanhuang MUD" (炎黄 MUD), npc_manager.py. Only public zh MUD+LLM signal found; single-NPC, no GitHub. Worth contacting the author if multi-agent MUD work proceeds.

Recommendations + Open items updated to gate the AgentScope spike behind the upcoming [[substrate-evennia-multi-agent]] exploration. Note's overall status stays converging; the L2 lean (ElizaOS-core + OpenClaw + Hermes) is unchanged.

Why this and not a new note

Adding to the existing prior-art-l2-l3.md rather than spawning a parallel prior-art-l2-l3-cn.md because (a) the L2/L3 categorization is the same, (b) the dataset is the same lineage (community/research work mapped to our needs), and (c) splitting by geography would bias future readers' mental indexing. The supplement is dated + headed for transparency about when it was added.

Test plan

• Review tables — L2 row reads consistently with the others; L3 rows ditto.
• All new URLs render correctly in GitHub markdown.
• Cross-link [[substrate-evennia-multi-agent]] resolves once the companion PR lands.
• Verify harness-ci signal (will run as path-filtered; should be green — no harness/ touched).

🤖 Generated with Claude Code