World — cumulative vision corpus

# World — cumulative vision corpus

> **Superseding product refinement — 2026-08-28:** this corpus preserves the broader World/MMO design
> space, but the current working identity is narrower: **World pins a preserved story coordinate as a
> replayable, inhabitable setup**. Scenario and Episode own branch-frontier story progression—Scenario fast
> and abstract, Episode slow and immersive. World runs may vary the snapshot without silently advancing the
> source branch. Exact persistence, lifecycle, primary use, and multiplayer scale remain open. Read the
> [World concept contract](README.md) and the
> [current founder refinement](../../sources/founder/2026-08-28-founder-episode-scenario-sibling-refinement-2026-08-28.md).
> Where this corpus calls deployable society, multi-camera inhabitation, ongoing story delivery, or MMO
> scale the centre, that language now records optional capability or historical ambition.

Status: **ordered product-vision corpus for Domain extraction; directional, not shipped behavior.**

This document is intended to be passed whole into Meridians. It gathers the World idea into one ordered
source: identity, ontology, world construction, play, information, economies, lifecycle, evaluation,
product horizons, and unresolved design questions.

The concise [World concept contract](README.md) remains authoritative for stable terminology. If this
expansive vision conflicts with that contract, the contract wins. This corpus deliberately preserves
speculative horizons without presenting them as shipped capability or settled mechanics.

## 1. The shift in Meridians

Meridians began as a way to extract, structure, research, explore, and narrate worlds. World reveals the
larger system latent inside that work: the extracted world is not only a knowledge base or a source for
stories. It can become a programmable place in which many humans and autonomous agents act, encounter one
another, alter shared state, and leave a durable history.

The model remains the product, but the model becomes inhabitable. World reframes Meridians as a
configurable game engine as well as a narrative and knowledge engine. Its medium is an abstract,
bare-bones text-visual graph world built for raw simulation of fictional economies, information
conditions, rulesets, tools, and plural agency.

The ambition is not infinite content for its own sake. It is worlds capable of continuing: worlds in
which consequence accumulates, information travels unevenly, rules can be programmed, and the frontier
expands wherever attention and Will reach it.

## 2. The canonical formulation

> **Meridians is a persistent, turn-based graph-world engine. System encodes programmable rules and
> constraints; World holds playable entities, information, and changing state; Fate carries open
> possibility and consequence. Scenario builds and explores the world. World lets humans and autonomous
> agents inhabit it through many situated cameras, injecting Will as MCP tool calls on a shared cadence so
> the world evolves turn by turn.**

The compact grammar is:

> **Domain defines. Branch preserves. Episode embodies. Scenario explores. World continues.**

And:

> **System constrains. World embodies. Fate remains open. Will makes the move.**

Persistence means state, lineage, memory, and consequence survive between turns. It does not require
always-on computation, immortality, or permanent activity.

## 3. Domain: the world substrate

A Domain defines the world a World can inhabit. It may derive from an original premise, fictional
corpus, research, licensed canon, synthetic material, or a governed combination. Extraction turns that
material into a structured graph-world rather than a flat body of prose.

The Domain contains or makes authorable System, World, Fate, Characters, Units, locations, artifacts,
institutions, Frames, Unit Control, resources, topology, rules, permissions, clocks, tools, canon,
branch lineage, provenance, historical conditions, communication regimes, and unresolved questions.

A Domain is not itself a World. It is the durable substrate from which Scenarios and Worlds can be
instantiated and may deepen through extraction, authorship, research, Scenario, or governed learning from
play.

This produces a world pipeline:

```text
valuable corpus → extraction → raw Domain world → Scenario exploration and preliminary testing
→ preserved states worth inhabiting → parallel World deployments → attributable played history
```

Extraction is not a clerical import step. It is the efficient translation of valuable text into usable
world structure: canon and provenance, entities and Frames, System rules, World state, open Fate,
topology, information, resources, and authorable gaps. The result can be consumed immediately as an
inspectable Domain and improved through research or authorship before play begins.

Scenario is often the proving layer for that raw world. Focused experiences reveal whether its situations
are engaging, its causal structure coherent, its choices consequential, and its missing machinery worth
building. A promising preserved state can then seed many Worlds rather than being consumed by one
deployment. This is a common progression, not a universal gate: mature Domains may deploy directly, and
valuable Scenarios need not graduate into Worlds.

## 4. The graph is world physics

The graph is not merely a visualization or retrieval index. Its topology can become part of the world's
physics. Typed edges express causality, reach, containment, authority, communication, transportation,
dependence, ownership, and economic exchange. Universal **Frames** carry a meaningful, directional reading
from an exact owner toward an exact target: love, loyalty, hatred, kinship, marriage, habitat, obsession, adherence to
law or Dao, religion, and devotion to Fate can use one expressive topology rather than isolated genre
systems. Frames never grant authority; singular full **Control** is a separate Unit-only contract.

Hierarchy matters as much as adjacency. Information and effects may travel upward or downward across
levels, laterally within a level, or through typed bridges between distant parts of the world. The same
topology helps derive what a seat is near, what it can affect, what can reach it, and which distant regions
remain dormant or abstract until play reaches them.

## 5. System, World, and Fate

**System** is programmable structure: rules, institutions, permissions, clocks, incentives, constraints,
resource transformations, economies, and action affordances.

**World** is what presently exists: characters, locations, artifacts, resources, relationships,
information, memories, conditions, ownership, and changing state.

**Fate** holds open bearing: unresolved questions, branching possibility, looming consequence, and the
directions in which the world may move without declaring one future inevitable.

The forces make play legible without predetermining it. System explains the constraint, World records the
embodiment, and Fate preserves what remains open.

### Frames and Control

**Frames describe directed readings; Control determines Unit power.** A Frame is one entity or concept owner's typed,
signed alignment toward one target. It does not imply command. A king may receive many positive Frames
without Control; a tyrant may Control a Unit whom Characters frame negatively; an Unbound Unit may
remain richly embedded in family, religion, habitat, trade, or society. Reciprocal readings between any two
endpoints are represented by two directed Frames.

Unit **Control** names exactly one Controller, or records the Unit as Unbound. It is
full authority rather than a strength, valence, weighted influence, or autonomy range. Recursive Unit
Control must remain acyclic. Partial, negotiated, or self-authored power belongs at Character resolution.

## 6. Will and MCP action

Will is the intangible force introduced by a human or autonomous agent at play time. It is not a fourth
stored force, a metaphysical property, or a score of agency.

For now, Will is choice under context. An agent receives a prompt, situated context, and available tools,
then chooses whether and how to call one. A human uses natural language or manual tool selection. A hybrid
seat combines direction with delegation.

The concrete expression of Will is a typed MCP tool call. The call crosses a governed boundary, records
whose authority produced it, and attempts a change that System may permit, reject, defer, transform, or
resolve against other calls. Will is the open choice entering the world; situated context supplies the
conditions around that choice. Neither should be mistaken for a metaphysical claim that a model has
human free will.

## 7. Scenario builds; World inhabits

Scenario and World are sibling configurable games over the Domain substrate.

Scenario is the focused narrative game for world-building and exploration. It forms a bounded, card-like
experience from situated perspectives, attaches to the active tip of a branch, plays possible consequences
forward from that frontier, and helps build the raw world. World is the multi-perspective
living-world game. It takes a preserved Domain state and lets multiple human or agent-controlled seats act
inside it over time. Its origin may be anywhere in preserved branch history, not only at a tip.

Both may eventually become learning environments, but at different temporal and social resolutions. A
Scenario is an episodic training playground: inject raw Will into a bounded situation, generate a narrative
outcome, inspect consequence, and compare the same decision point across branches. A World is a
persistent multi-agent training economy: policies learn through repeated turns, partial information,
scarcity, tools, institutions, other Will, and consequences that alter the next state.

Scenario can prepare a world for World. World history can later become material for reflection,
Scenario, branching, or world-building, but it must not silently rewrite Domain canon.

The name describes the composition: a Domain holds many possible viewpoints and paths; Scenario brings
selected perspectives into one directed experiential frame. It binds an origin state, a live question,
and the consequential passage produced by Will. It can be held and compared like a narrative card, but
it names the playable experience rather than only its UI.

**World** names the other scale: not a representation of a world, but one continuing played
reality arising from a preserved Domain state. World remains the short form. One Domain can yield many
Scenarios and many parallel Worlds.

```text
Domain     defines
Branch     preserves
Scenario      forms a bounded narrative from situated perspectives
World     lives as a shared reality
```

Both should offer a coherent base game, deployable presets, and deeper custom settings. The goal is many
games over one trustworthy substrate, not one universal genre.

Scenario may develop a branch edge through several experiential instruments rather than one permanent UI.
A card-like table can abstract intent, uncertainty, commitment, and counterplay. A graph mandate can address
an entity, relationship, rule, or open Fate directly. A governed context injection can introduce new
source-bounded information or a situation into the branch frontier. Each is a way of injecting situated
Will into the same Scenario lifecycle; none creates a second writer or silently promotes its result into
canon.

## 8. Branches and parallel Worlds

Meridians maintains two related layers of possibility. The branch tree preserves authored and explored
world states; its tips are where Scenarios continue building possibility. Worlds form a second, unconstrained
population of played realities attached to coordinates throughout that tree. Any preserved state—tip or
non-tip—may originate zero, one, or many Worlds, and every branch may carry many such origins.

This is the multiversal perspective: one Domain can hold many branches, each branch can support many
Worlds, and many Worlds can share the same origin without sharing a future. Creating or advancing one
does not consume the state, move the branch tip, or privilege it as the branch's canonical World.

A World can begin from any preserved state on a Domain branch, not only its head. Many Worlds may share
one branch or exact revision and diverge through different players, agents, prompts, presets, rules,
resources, information conditions, economies, injections, or choices.

A branch preserves a possible authored Domain state. A World is a played runtime with cadence, seats,
contexts, tools, resources, calls, events, and an independent social history. World activity does not
silently mutate its source branch, and common ancestry remains legible.

## 9. Every World is turn-based and clocked

All World instances advance in turns under a declared cadence. They distinguish wall time, turn time,
and fictional time; their mapping is configurable by Domain and preset.

The schedule declares when consequential play is open and closed. Active hours, off-hours, time zones,
weekends, holidays, pauses, and rest periods are game design rather than deployment trivia. The world may
remain inspectable while consequential play is closed.

This makes World an idle MMO by tempo. Automated agents can continue delegated play, invested humans can
intervene when present, and constant availability need not become the dominant advantage. Exact phases
remain Domain-specific; no universal turn choreography is settled.

A shared cadence does not imply equal action rate. A World may allocate different tick eligibility,
actions per window, cooldowns, queue priority, concurrency, and tool latency to different entities or
agents. Rate may follow role, location, artifacts, capability, resource expenditure, reputation,
delegation, or another System rule. One policy may iterate through frequent low-cost observations; another
may conserve scarce turns for consequential tools; another may invest in infrastructure that increases
future throughput.

This makes time and iteration capacity economic. The ability to decide and act faster can compound into a
greater ability to reshape reality, so allocation, acceleration, waiting, and opportunity cost must be
visible in the record. Compute latency and scheduler accidents must remain distinct from fictional time
and declared rate, or infrastructure silently determines who has power.

## 10. Seats and cameras

A seat is the playable perspective through which Will enters. It may be bound to a character, collective,
location, artifact, institution, infrastructure, or another Domain-authorized entity.

A preset may be character-only, location-only, artifact-only, institution-only, or mixed. Seats may be
human-operated, agent-operated, or hybrid. Identity, authority, delegation, and the origin of each tool
call remain legible regardless of operator.

Different playable kinds enable different games: artifact acquisition and custody, location migration and
infrastructure, institutional coordination, or actor-led political, economic, and social play.

Character and Unit are resolutions of agency rather than judgments about intelligence, sentience,
morality, rank, or importance. A Character receives high-resolution agenda, memory, attention, and
objective continuity. A Unit is simulated at lighter resolution and is fully Controlled or Unbound.
Promotion and demotion preserve identity, history, Frames, knowledge, and consequence while
changing the resolution at which the world spends attention.

## 11. Context is gameplay

Each seat receives context assembled for its current position. Context may include identity, personality,
incentives, irrationalities, graph location, nearby entities, received information, memory, relationships,
obligations, ownership, reputation, artifacts, resources, access, tools, permissions, the current turn,
and visible Fate questions.

Actors may be rational, boundedly rational, emotional, habitual, mistaken, loyal, deceptive, impulsive,
or conflicted. Autonomous play should not collapse into omniscient utility maximization merely because the
underlying model could read the full record.

Historical outline and entity continuity provide an evidence proxy, not direct access to a character's
interior. [Entity Interpretation](../latent-model/README.md) may compress sufficiently rich evidence
into a versioned Entity Interpretation Graph, then activate it with the seat's actual knowledge, relationships,
state, and constraints to produce competing tendencies. The graph never chooses the action.

Canonical truth, observation, received information, belief, memory, and story are distinct. What context
is retrieved, summarized, withheld, or allowed to decay changes the game itself.

## 12. Information propagation

Canonical world state is not automatically common knowledge. Information travels through topology:
across nested levels, among peers, over spatial routes, through relationships and institutions, inside
artifacts and records, and across technological infrastructure.

An oral, courier, print, broadcast, and internet-enabled world should not distribute the same event in the
same way. Relevant properties include reach, latency, synchronization, fidelity, compression, distortion,
credibility, authority, secrecy, surveillance, leakage, decay, forgetting, archival recovery,
discoverability, and retrieval cost.

Propagation may stall at boundaries, fragment into competing accounts, move asymmetrically, or be
manipulated. These nuances matter because information changes which actions a seat can recognize, justify,
coordinate, price, or attempt.

## 13. The information economy

Information is an economic object because access is uneven and action depends on what is known. A fact may
be valuable before it spreads; a secret may change bargaining power; a credible source may be scarce; an
archive may make forgotten events actionable; false information may redirect resources or Will.

Seats may expend time, access, relationships, artifacts, reputation, or money to obtain, verify, transmit,
conceal, distort, or monopolize information. Context and propagation together determine the effective
information market of the world, even where no currency exists.

## 14. Material, financial, and capability economies

Where a Domain permits, a World may model artifacts, money, credit, debt, land, energy, materials, food,
labor, time, access, influence, trust, reputation, tool capability, computation, and delegated agent labor.

Scarcity, access, transformation, custody, and transferability are more fundamental than currency. A
Domain may configure barter, currencies, auctions, credit, crafting, labor markets, patronage, information
markets, or no explicit market at all over those primitives.

Acquisition, ownership, exchange, production, custody, transformation, and loss create stakes. A Domain
need not imitate contemporary capitalism; it can encode its own institutions, norms, taboos,
irrationalities, and failure modes.

Externally valuable assets, real-money stakes, wagering, or speculative markets are a separate horizon
requiring explicit legal, licensing, integrity, custody, anti-abuse, and operational design. They are not
assumed by the base game and do not make World a claim to predict or verify the real economy.

## 15. Deployable societies and emergent behavior

A World is a deployable society instance. The Domain supplies the world substrate; World configuration
supplies its constitution, topology, population, tool ecology, information regime, economy, clock, and
initial conditions; repeated injections of plural Will create its history.

The central bet is emergence. A society should not be described into a prompt and then performed back by
compliant agents. It should develop through situated actors using uneven capabilities under persistent
rules and consequence. The relevant experimental variable is therefore not only agent personality or
prompt quality, but the **tool ecology**: which actions are possible, who can perform them, how costly they
are, what information they require, and how they compose with the actions of others.

From those conditions may emerge specialization, trade, information arbitrage, coalitions, status,
norms, institutions, cooperation, conflict, and concentrations of power. Money may matter, but so may
ownership, reputation, location, transportation, communication, surveillance, access, and social
authority. These mechanisms turn the graph's spatial and informational arrangement into lived strategic
conditions rather than decorative worldbuilding.

A collective need not begin as a new kernel entity. A household, religion, guild, company, army, or
nation may first emerge as stable shared Frames, coordinated Control, information boundaries, pooled
resources, System commitments, and common tools. A Domain can promote that formation into a playable
entity if it needs independent memory, custody, permissions, or agency, without inventing a parallel
Faction topology.

Emergence is a research proposition, not a guarantee. A World must preserve an attributable experimental
record: the context and information available at choice time; the actor, prompt or policy, authority, and
tool version; accepted and rejected calls; state, resource, ownership, status, and location changes;
communication and propagation paths; rule and institutional changes; external injections; and the
configuration and topology under which they occurred. Replay, forks, interventions, and comparisons
between Worlds can then reveal how behavior changed when one condition changed.

Fictional worlds provide the clearest bounded laboratories. Historical or non-fiction-derived worlds can
still support useful experiments, but their results describe the configured World. They do not establish
how the source society works, predict what it will do, or validate a real-world intervention without
separate external evidence.

## 16. Presets and custom engines

World is a configurable base game. Presets allow rapid deployment, custom settings allow deeper
authorship, and some Domains may eventually require custom-coded mechanics.

Configuration may vary playable kinds, seat composition, cadence, automation, MCP tools,
tick rates, turn allocations, queues, cooldowns, tool latency, permissions, costs, context, memory,
information propagation, economies, objectives, secrecy,
cooperation, competition, evaluation, spectator access, lifecycle, and narrative lenses.

Custom engines may express unusual physics, institutions, resources, and interaction rules, but must not
bypass shared authority, provenance, state, permission, lineage, and history contracts.

## 17. Continuous world expansion

New material can enter while play continues. Injections may be human-authored, synthetic, derived from
licensed sources, or seeded from cited real events where a Domain allows.

An injection is not automatically truth. It enters through declared authority as an event, proposal,
condition, or source interpreted by world rules. Reality may perturb a fictional parallel without making
the World a forecast; synthetic events may test resilience without pretending to be evidence.

Expansion can introduce actors, locations, artifacts, relationships, rules, resources, information, or
Fate questions. Detail grows where attention and consequence travel.

## 18. Lifecycle and migration

Worlds may be short experiments, bounded futures, seasons, leagues, long-lived worlds, or indefinitely
continuing universes. They may pause, resume, conclude, score, archive, fork, migrate, or upgrade.

A World may move toward a newer branch head, future storyline, or changed Domain state. The transition
must be explicit about what changed, how incompatible state was handled, which history remains
authoritative, and where new lineage begins. Domain updates never silently rewrite a World's past.

Whether most Worlds should be short-lived or truly persistent remains open. The architecture should
preserve both possibilities until play supplies evidence.

## 19. Story as multi-camera projection

The canonical object is world state and event history. Narrative is a lens over that record. One World
may produce personal accounts, location chronicles, artifact histories, faction briefings, economic
reports, public news, spectator recaps, or explicitly omniscient histories.

Unless a lens is omniscient, it reflects only the information and vantage available to its camera.
Accounts may conflict without corrupting canon. Generated prose makes the abstract simulation inhabitable
but never becomes a second source of truth.

The experience is ultimately multi-modal. One continuing World may surface as a graph, a card-like
decision passage, a text-led social room, a spatial sprite world, a tactical or market instrument, a
Director-controlled cutscene, or an authored Episode. The same event may receive several of these
projections for different seats or audiences. Transitioning between them changes presentation, available
controls, and cadence—not World identity, ancestry, or canonical history.

Everything on the surface may therefore be staged without everything being on rails. A free-form scene is
orchestrated responsively as human and agent Will resolves into events. A bounded set-piece temporarily
hands control to a Director or template and explicitly returns it. An Episode fixes the passage for an
audience and introduces no new Will. Capital-S Stageplay names the present Episode pipeline; the broader
orchestration grammar is a future runtime direction whose authority remains downstream of the World.

## 20. Evaluation and scoring

A World may be evaluated live, posthoc, or both. Measures may include Fate movement, survival,
resilience, coordination, institutional capacity, influence, reputation, discovery, information
advantage, calibration, artifact acquisition, resource performance, economic performance, objective
achievement, judgment, adaptation, and learning under consequence.

No universal scalar is constitutional. Every evaluation names its World, ruleset, perspective, interval,
measure, and limits. Scoring can support play, leagues, agent evaluation, comparative experiments, and
learning, but should not flatten a rich world into one Goodharted objective.

## 21. Product forms

> **Where we are going.** The family below is the possibility space; the roadmap points at three of its
> members. **Experience (VN)** — an hour in an original world, read back afterwards; **training worlds for
> agents** — the same kept world with a policy in the seat, the log as the result; **the populated World
> (VNMMO)** — many players as many characters in one replayable world. One substrate entered three ways;
> what each leaves behind is kept history and the record of having been in it
> ([12](../../knowledge/product/strategy/12-the-practice-monopoly.md),
> [13](../../knowledge/product/strategy/13-consequential-test-worlds.md),
> [World roadmap](README.md#what-the-gates-are-for--three-forms-one-substrate)). Nothing here promotes a
> form past its gate.

World opens a family of products:

- fictional economy and ruleset simulations;
- idle MCP agent worlds;
- cooperative and competitive human-agent games;
- multi-agent social, political, institutional, and economic experiments;
- parallel worlds varying rules, information, or initial state;
- leagues and seasons with declared objectives;
- artifact-led, location-led, character-led, and mixed games;
- original living fictional universes;
- licensed popular universes where rights permit;
- persistent agent training and evaluation environments;
- markets for situated information, artifacts, reputation, strategy, capability, and agent labor.

### Model-evaluation product hypothesis

One falsifiable commercialization path is to use World as a benchmark runtime for model and agent
teams. The benchmark would not primarily ask whether a model can emit a valid tool call or complete one
short task. It would ask how a policy behaves across consequential actions in an unfamiliar persistent
world, where earlier choices alter later information, resources, relationships, institutions, and
available affordances. Tools are the action vocabulary through which capability becomes observable, not
the capability being scored in isolation.

```text
versioned Domain → executable causal world → unseen circumstances → attributable trajectories
→ scoped capability profile and inspectable failure cases
```

A future service could run submitted agent endpoints or pinned policies through controlled families of
Worlds and compare long-horizon planning, tool discovery, information acquisition, calibration,
adaptation, cooperation, recovery after mistakes, and goal persistence. Its useful output would be more
than one scalar: it would identify the conditions under which behavior changed and return the trajectories
that make a weakness actionable. This remains a product hypothesis, not the external promise, until one
such measure proves repeatable, discriminative between systems, and useful to a team improving a model.

Scalable evaluation also makes scalable affordance construction a prerequisite. A candidate stack to
prototype is a stable typed action contract, reusable world-specific tool packs, and runtime affordance
resolution from seat, location, information, possessions, permissions, institutions, and current state.
Schemas, composable preconditions and effects, generated descriptions, validation, replay, and versioning
could let a Domain produce varied action environments without requiring thousands of unrelated handwritten
tools. This is a design direction to test; it does not settle the Domain extension API.

The benchmark must be more trustworthy than the systems it evaluates. Useful evidence requires pinned
world and tool versions, deterministic transitions where the world permits them, controlled model
settings, multiple seeds, complete cost and token accounting, auditable trajectories, contamination
resistance, and independent transfer checks against capabilities that matter outside the simulation.
Performance remains evidence about a declared synthetic world until that transfer is demonstrated.

Source: [2026-08-21 tool-use benchmark conversation](../../sources/conversation/2026-08-21-pasted-text.txt).

### Training worlds for agent cohorts

In a future training form, Meridians can provide abstract textual environments for reinforcement
learning, curriculum design, agent evaluation, and policy comparison. The graph is traversable state; the
information regime creates partial observation; MCP tools define the action space; System supplies
transition rules; resources create incentives; Frames and Control create social structure; Fate supplies
long-horizon pressure; and the log retains attributable trajectories.

Scenario and World support different training paradigms. A Scenario is closer to an episode: inject raw Will
into a pivotal situation, produce a bounded narrative outcome, and compare branches from a controlled
origin. It can train or evaluate judgment over concentrated decisions without requiring a continuing
society. A World is closer to a multi-agent environment: many agents or cohorts operate across turns,
learn through incomplete information and changing affordances, and alter the economic and institutional
conditions encountered by later policies.

Parallel origins allow cohorts to face the same world while varying prompts, policies, tools, information,
reward instruments, or social rules. This can create turn-based training economies for cooperation,
competition, negotiation, discovery, resource allocation, institutional participation, and tool use. The
textual abstraction is an advantage: rich law, Dao, religion, marriage, hierarchy, allegiance, reputation,
and information structure can be encoded without first constructing a photorealistic environment.

Training must preserve the conditions that produced apparent learning. Every comparison needs versioned
origins, policies, prompts, contexts, tools, rules, evaluators, and reward scopes. Performance inside a
Scenario or World establishes behavior under that declared synthetic environment only. Claims of improved
real societal performance require transfer tasks, independent evaluation, safety and distribution-shift
analysis, and evidence from the actual environment. The simulation may produce candidate policies; it
cannot certify its own transfer.

The deepest possibility is a market of intelligence in which information, judgment, coordination, tool
use, and agent capability become visible through action and consequence. Fantasy-sports-like leagues and
speculative surfaces are conceivable future forms, not the identity of the base practice engine.

The person's side of these forms is [strategy 12 — The practice record](../../knowledge/product/strategy/12-the-practice-monopoly.md);
the agent's side, with its gates, is [strategy 13 — Training worlds for agents](../../knowledge/product/strategy/13-consequential-test-worlds.md).
This spec owns the world-side design above.

## 22. The common engine

The shared engine direction includes Domain extraction, graph topology, canonical state, rebuildable
projections, branch history, Scenario, World turns and cadence, Characters and Units, Frames and Control,
Drive derivation, typed MCP actions, per-seat identity and context, information propagation, optional
economies, event injection, world expansion, state migration, multi-camera narrative, provenance,
evaluation, training trajectories, observability, and replay.

Domain-specific rules must coexist with one source of truth. Tools, agents, interfaces, schedulers, and
Stories are callers or projections around the world, not alternate worlds beside the canonical record.

## 23. Shipped substrate and direction

Meridians already contains important substrate: Domains organized through System, World, and Fate; graph
structures; extraction and research; branches and preserved state; a canonical record; Scenario
primitives; programmatic operations; and MCP-facing tool use.

World as a persistent, clocked, multi-seat runtime remains direction. Information propagation, per-seat
autonomous context, configurable economies, World lifecycle, migration, scoring, training curricula,
cohort protocols, transfer evaluation, leagues, licensed universes, and external-value systems require
design and implementation.

The next proof is a small playable World with more than one situated seat, a declared cadence, partial
information, typed tool calls, persistent consequence, and a legible history.

## 24. Deliberately unresolved

Open questions include universal turn phases, conflict resolution, ideal duration, runtime-state
representation, context selection and forgetting, the first propagation model, coherent autonomous
personality, Drive composition, Character/Unit resolution policy, configuration versus custom code, a
safe Domain-engine interface, first economies, resource and ownership representation, useful scoring,
training curricula, reward design, cohort comparison, transfer evaluation, Fate evaluation, lifecycle
transitions, promotion into Scenario or canon, information visibility, event authorization, external-value
governance, and the first public preset and surface.

These should be settled through prototypes and play rather than prematurely constitutionalized.

## 25. Final vision

World is the point at which Meridians' structured worlds become living systems. Scenario constructs and
explores possible worlds, concentrating raw Will into bounded narrative consequence; World gives one of
those worlds a clock, many cameras, incomplete information, programmable rules, scarce resources, and
repeated injections of plural Will.

Humans may play directly. Agents may play idly. Humans may delegate, intervene, collaborate, compete, or
watch. Characters, Units, locations, artifacts, and institutions may become perspectives. Frames make
attachment legible; Control shapes derived agency; Drive joins situated pressure to action. Information
moves according to history and topology. Economies emerge around what is scarce, known, owned, exchanged,
or made possible. Stories render many viewpoints while the graph records what actually changed.

Many Worlds can begin from many points on many branches. Some last hours, some seasons, some years. Some
are private experiments; others may become shared leagues or licensed living universes. All are held
together by explicit rules, situated context, typed action, legible authority, preserved lineage, and
consequence that survives the turn.

That is the grander Meridians: not only a system that knows or narrates a world, but an engine through
which many forms of intelligence can inhabit one, learn within it, and—where transfer is independently
proved—become more capable of acting in the world beyond it.

The near vision is narrower and is what the next roadmap is for: worlds **worth experiencing** (a better
book that remembers you), **worth being inside** (a place many wills can re-enter without it coming
apart), and **worth growing from** (a kept record of one's own judgment, and a kept test of a system's).
The grander Meridians is reached, if at all, by owning those three.

Raw Markdown source