docs(exploration): substrate option — multi-agent text MUD (Evennia + N×LLM NPCs)

[cyber-ayi]

Summary

New exploration note evaluating whether the ladder (adr-0007) should replace its deferred AI Town slot with a self-hosted text MUD (Evennia, possibly DikuMUD-area-fed) populated by N≥3 autonomous LLM-driven NPCs + one operator + one designated "main agent."

Surfaces three core tensions, gates resolution on three concrete spikes (S1/S2/S3) and an operator UX ack. Status remains open pending operator co-presence ack.

Why this is more than a no-op

Per the companion PR's zh prior-art supplement, GenerativeAgentsCN empirically validates N=25 LLM agents running on local Ollama + Qwen3-4B. Combined with telnet/Evennia integration cost being lower than AI Town's Convex+TS stack and project-stack fit (Python+Ollama already in tree), Option B strictly dominates AI Town on the dimensions AI Town was placed on the ladder for.

Why this isn't a P0/P1 replacement

• Melvor P0 lock (adr-0003 / adr-0007) holds — multi-agent + non-determinism before the apparatus is proven violates the clean-room principle.
• Stardew P1 is on the ladder for operator co-presence warmth (named NPCs, calendar, gifts, graphical) — Option B doesn't replicate that.

The three tensions

1. Dyad-vs-multi-agent attention split — mitigation: foreground/background architectural asymmetry (operator + main agent fully-modeled; NPCs small local + short context + not in main agent's reflection loop). Load-bearing constraint — without it, Option B drifts into "another Generative Agents + Concordia."
2. Setting choice — wuxia gives a 30-yr zh asset trove but imports a 升级打怪 culture that risks adr-0004; cannot resolve until S2 + operator setting ack.
3. Determinism is unprecedented at this combination — Paracosm / Miniverse close on a subset, neither MUD-shaped. Either pay the cost (S3) or accept non-CI-replayable runs.

Gating spikes (must all pass)

ID	Spike	Output
S1	Fork GenerativeAgentsCN; N=25 with Qwen3-4B on operator HW; measure tokens/tick.	cost-ceiling number
S2	Import DikuMUD tbamud area into Evennia + skim wuxia-MUD lib (pkuxkx wiki / mudcore).	content-track decision
S3	Single-node Ollama + fixed seed + greedy + logical tick + content-addressed LLM-response cache. Byte-equal replay of 50-tick scenario.	determinism feasibility
S0	(after S1–S3) AgentScope vs ElizaOS-core comms bench-off.	runtime within Option B

Companion PR

#30 — adds the zh supplement to exploration/prior-art-l2-l3.md that this note references via [[prior-art-l2-l3]]. The two PRs are intentionally separate: the prior-art update is independently useful regardless of how Option B resolves.

Test plan

• Frontmatter validates (status: open, related-adrs cited, resolves-to blank).
• All [[name]] links resolve to existing notes or are intentional forward references.
• All external URLs render and are reachable.
• Read the "Three core tensions" section critically — these are the load-bearing constraints; flag anything that feels weak.
• harness-ci signal (path-filtered; no harness/ touched, but runs on every PR per recent change).

🤖 Generated with Claude Code