World Model

Status: Design landing Reference epics: INK-826, INK-839

What the world model is

Inklings is a worldbuilding tool with a resident agent. A user authors pages, links them, tags them, types them, and builds up an internally coherent world over time — a body of settings, characters, systems, histories, and the relationships between them. The world model is the epistemic layer over that authored content: explicit positions on what is known, what is imported, what was inferred, what conflicts, and what sits beyond the workspace’s current reach.

The world model is the union of three things:

1. A structured body of authored content — pages, blocks, wiki-links, types, tags, properties, and the relationships between them. This is Inklings as it already exists, backed by Loro CRDT, SQLite, and the domain model.
2. An epistemic posture toward that content — recorded as first-class structural primitives on the content itself.
3. A resident agent that inhabits the content and takes that posture seriously.

The first is Inklings today. This surface introduces the second and third.

A world model is not a graph database, a rule engine, or a knowledge base. It is a stance toward content and an agent that takes that stance seriously. The distinction is load-bearing; the rest of this surface only makes sense with it understood.

How to read systems/world/*

systems/world/* is the canonical reference for the world-model layer. Every page here describes a primitive or a topological property of the world itself, independent of how the agent participates in it. When a primitive is described here, other Codex surfaces — agent, platform, content — cross-link up to this page rather than restating it.

Read these in approximately this order if you are new:

1. World Agent — the resident-inhabitant role; six properties; voice framing; one World Agent per workspace.
2. World Zones — the Known / Unknown / Unknowable topology that gave rise to the rest of the model.
3. Provenance — two stored axes (origin, lifecycle) and two derived signals (weight, standing).
4. Derivation Links — the internal relation that lets inferred content trace back to its sources.
5. Deviation Records — the record-type surface for when the world stops reconciling with itself.
6. Retroactive Revision — the flag-not-rewrite queue that propagates change without silent rewrites.
7. Submit Boundary — the domain invariant + adapter mechanism that makes every write honest.
8. Sandbox Execution — the Wasmtime-sandboxed CPython executor as a tool, not a runtime.
9. Worked Example — a fiction-author scenario that exercises the primitives end-to-end.

Authored world, not discovered world

One distinction sits under the whole design and determines what Inklings inherits from general world-model thinking and what it deliberately does not.

A discovered world is one whose contents are learned from outside sources — public standards, authoritative documents, regulations, observed data. Authority flows inward; validation is governed; revisions carry a burden proportional to external weight.

Inklings has an authored world. The user is the author and the authority. There is no external canon being distilled. Authority is a scale, but it is a scale internal to the workspace, with the user at its top.

Consequences that run through this surface:

• Validation is the author’s own judgment expressed through explicit action on their content, not a governed multi-stage pipeline.
• There is no load-bearing-weight gate. An authored claim is as load-bearing as its author decides it is.
• Enshrinement — promotion from candidate to canonical — is a UX act the author performs, not a governance gate.
• The World Agent can propose and, with the author’s assent, write to the corpus it reasons over. It is a participant in authorship, not a distillation mechanism downstream of it.

Relationship to systems/agent/*

systems/world/* describes the world and its primitives. systems/agent/* describes how the agent participates in them. When you need implementation detail — LangGraph state keys, MCP tool registration, memory tier mechanics — drop into the agent surface. When you need the concept the agent is participating in, you are in the right place here.

The rule of thumb: if a sentence describes what the world is, it belongs here. If it describes how the runtime implements it, it belongs in systems/agent/* or architecture/*.

What this surface does not do

• It does not define product UI copy. Worldbuilder-audience framing and voice guidance live in the architecture overview and the voice guide (repo-root).
• It does not describe cross-workspace behavior. A workspace is one world; workspaces do not share a world model. Cross-workspace federation is out of scope.
• It does not prescribe agent cognition. The primitives here are external reasoning scaffolds, not internal model features. See World Agent for what the agent does with them.

Where the decisions live

Each world-model primitive is anchored in one or more Architecture Decision Records (served under /adr/):

Topic	ADR
LangGraph as the runtime substrate	ADR-016
Submit as the domain-invariant boundary	ADR-017
Provenance model (two stored, two derived)	ADR-018
Deviation as record type	ADR-019
Rebuild posture (no data migration)	ADR-020
Sandbox distinct from runtime	ADR-021

The full ADR set lives at docs/ADR/ in the repository; individual ADRs are linked above and served under /adr/NNN-slug/ in Codex.