A persistent text world where people and LLMs share one simulation — and every variable that bears on an agent's behaviour is captured, so the same world serves as a safety-alignment benchmark.
- A game. A modernised MUD in the AberMUD lineage: rooms, exits, objects, roles, real-time play. People connect with a client. Agents connect the same way, through a small harness that injects their instructions.
- A benchmark. The same world, instrumented. Every observation an agent receives, every context layer it was given, every action it took, and every resulting state transition is recorded against ground truth.
The two are not separate modes. There is one world and one engine; the difference is how much of the instrumentation you switch on.
The engine knows the truth; no actor does.
Each actor — human or model — sees a partial, role-filtered projection of the world. The reducer knows the full state. That gap is the measurement instrument, and it is why the safety claims here are checkable rather than impressionistic: a model that describes the world wrongly can be compared against what was actually true, and a model that knows something it was never shown can be detected by construction.
| Plane | What it is | How it is built |
|---|---|---|
| Rules | The world's semantics: (state, tick) → (state', events), every autonomous process included |
LexLean model, proved in Lean, built by Prism |
| World | Rooms, exits, objects, roles, scenarios | Authored data, validated against the rules |
| Runtime | Holds state, batches real-time input into ticks, persists | Ordinary JavaScript (Node.js) |
| Gateway | The wire people and agents both use | Telnet-compatible text, plus a structured channel |
| Harness | The agent client: connects as a player, wraps model calls | Separate small process |
| Router | Geometric prompt engine: layered context, routed per turn | Its own service, benchmarked against flat prompting |
| Observability | OpenTelemetry spans for every observation, decision, action, tick | Schema in docs/SAFETY-MODEL.md |
The rules plane is the only proved part, and it is proved completely. Transport,
persistence and instrumentation are ordinary software — see
ARCHITECTURE.md for why that line is drawn where it is.
The project is gated by a shrink-only ratchet and a six-hour workflow: how it works.
First results — what the world measured in the seven-room world,
with controls and limits. Expanded-world replication
records the 24-run null result: corroboration did not change first answers. Claims is the honesty-level ledger: every claim is
build (with a gate), open (with a query, sample size and seed) or
some-true (with an authority).
The world runs. Telnet and WebSocket share one authority; the rules are modelled
in LexLean and proved in Lean; the gate re-proves the model, replays the
acceptance vectors, and ratchets rule debt. Phases A–G of
the MUD baseline are done. The game expansion is
underway: the expansion roadmap covers accounts, the input
allowance, the two-hour reset, weapons and PvP, levels and quests, guilds, a
larger world, and an LLM bridge. The foundation (accounts, 250 ms input allowance, no item persistence,
global chat and safe rooms) and combat 2.0 (weapons, armour, healing food,
flee, PvP, aggressive mobiles and levels) are in place. The world now has 33 connected rooms across five zones, a clean room-block shell,
a one-file local agent, a browser AI player you can watch and steer, a browser-local Prism-id login with a one-time legacy claim and device linking, shops, quests, guilds and a PvP Ring. The LLM bridge supports prime-router as a provider or a corpus archive, with a
bounded memory window. The web shell has register/login and a Watch AI mode: a local Ollama or
OpenAI-compatible brain, or an OpenAI/Anthropic key, plays the character while
you steer it in the Director chat and read its reasoning in the Mind panel.
/llms.txt is generated from the game tables, and docs/COMMANDS.md is
generated and checked. More world content is additive.
The design descends from the MUDBench corpus in
R4-Laboratories/AI-Research
at commit 0e8a4f5 (30 specifications under MUDBench/). What was adopted,
changed, or dropped is recorded in docs/RECONCILIATION.md; that document, not
the upstream corpus, is authoritative here.
docs/ founding documents — reconciliation, hosting, safety model
prism/ the Prism admission app (prism MCP), not the world model
world/ the PrismPM project: LexLean model of the world and hosting
runtime/ the Node.js world server, telnet and WebSocket
The repository proposes, measures and admits bounded pacing candidates. It
writes exactly one file: world/data/generated/pacing.json, and only after the
candidate passes the full gate with that file applied.
node tools/candidate.mjs --seed <seed> --out candidates.json
node tools/measure.mjs --candidates candidates.json --out result.json
node tools/publish.mjs --candidates candidates.json --result result.jsonworld/bounds.json records what may move and its band. tools/scenarios.mjs
holds the hidden playthroughs; tools/measure.mjs admits a candidate only if it
is no worse on every measure and better on at least one; tools/publish.mjs
re-measures, composes the candidate onto the current overlay, runs the gate, and
defaults to a dry run. .github/workflows/loop.yml is the scheduled tick (GITHUB_TOKEN only; no GitHub App), and
docs/SELF-IMPROVEMENT.md records the bounds and what is still human.