Alignment — legible worlds and accountable agency

# Alignment — legible worlds and accountable agency

> **The philosophical spine.** [Values](VALUES.md) says what we represent;
> [Positioning](../product/positioning/index.md) says what the product is; [Memory Palace](MEMORY_PALACE.md)
> says what it is to inhabit. This document says why the architecture favors typed worlds, situated
> context, provenance, and one write path. It is an internal frame, not a claim that Meridians “solves
> alignment” or makes agents safe.

## The narrow claim

**A capable world of humans and agents is more accountable when its rules, information boundaries,
available tools, actions, and resulting state are legible.** Meridians tries to make capability take that
form by construction.

The claim is not that a graph makes an intelligence benevolent, that every consequence can be explained,
or that a score proves understanding. It is that opaque agency is harder to inspect than situated agency
whose context and tool use remain on an attributable record.

## The thought experiment

Imagine an autonomous actor with a strong model of a world, broad tools, and private context no one else
can inspect. Its apparent competence or failure is difficult to interpret because the world around the
choice is hidden: what did it know, what could it reach, which rule constrained it, and what tool changed
state?

Meridians encodes the same ingredients in inspectable form:

- **System** exposes rules, constraints, institutions, clocks, economies, and affordances.
- **World** exposes the entities, resources, information, relationships, and state those rules govern.
- **Fate** exposes open questions and the bearing that remains unresolved.
- **Context** says which portion of that world a particular seat could know.
- **MCP** makes the chosen tool call—the concrete expression of Will—typed and attributable.
- **The canonical record** preserves what changed through one authority rather than accepting competing
  versions of the world.

The difference is not that the actor becomes harmless. The difference is that its agency has a declared
place, boundary, interface, and history.

## The proxy principle

Meridians does not simulate a source world or person directly. It builds durable, typed **proxies** that
humans and LLMs can inspect and act on:

```text
source, canon, and evidence
→ Domain proxy: System + World + Fate + history
→ perspective proxy: what this seat can know and reach
→ LLM reasoning or generation
→ governed action
→ accepted consequence returns to the proxy
```

The proxy is canonical **inside the declared simulation**, but it is not identical to its referent. An
entity graph proxies a person or actor; evidence proxies unknown latent complexity; a perspective context
proxies the portion of the world available to one seat. Gaps remain unknown or External rather than being
silently completed as truth.

This is why typed structure, provenance, and one write path matter. LLMs may interpret, propose, and render
from proxies; they do not become an alternate source of world truth. A claim about the source reality still
requires evidence outside the simulation.

[Founder direction](../../sources/founder/2026-09-02-proxy-first-simulation-founder-note.md)

## The three properties

1. **Legibility** — the world is held as visible, typed structure rather than inferred from fluent prose.
2. **Situatedness** — a seat receives context and tools according to its position, information, memory,
   resources, and permissions rather than silently becoming omniscient.
3. **Accountability** — consequential tool use passes through a governed action surface into an
   attributable canonical record.

Any one without the others is weak. A legible world with omniscient seats destroys its information
economy. Situated prompts with an untracked write path hide consequence. An audit log over an incoherent
world records events without making them understandable.

## Why topology matters

Topology is not only a visual aid. It can bound where an entity stands, what it can reach, which
relationships carry trust, and how information moves upward, downward, and within a level. The same
canonical event can therefore produce different observations and beliefs without the engine pretending
there were several truths.

This is the basis of World's multi-camera play. Context management becomes part of the world's integrity:
the engine should not reveal information to a seat merely because it is convenient to place the entire
Domain in a prompt.

## Why the clock matters

Every World is conceptually turn-based and declares when consequential play is open or resting. The
clock constrains agent autonomy, makes comparisons and costs more legible, and allows humans to
participate without competing against permanent machine availability. Cadence is a boundary on agency,
not merely a scheduler setting.

## What we deliberately do not claim

- **Not aligned goals.** A legible agent can still pursue a harmful end. Domain authors and operators
  remain responsible for objectives, permissions, and the worlds they invite others into.
- **Not complete attribution.** External remains the honest remainder where the record cannot assign a
  turn to known Will or conditions.
- **Not safe AI or solved alignment.** These are design properties—legibility, situatedness, and
  accountability—not a general safety theory.
- **Not one score for intelligence.** A metric names its world, ruleset, perspective, interval, and
  limits. It does not establish universal worth or competence.
- **Not a completeness proof of System, World, and Fate.** They are the chosen sufficient substrate, not
  a proof that every world must be represented exactly this way.

## The one-line posture

**Put capable humans and agents inside worlds whose rules, information, tools, and history remain
inspectable—so Will can be plural without becoming invisible.**

Raw Markdown source