Skip to content

The present research

Begin with relation,
not the isolated self.

ARCANA is a computational architecture for asking how an artificial agent forms, carries experience, reasons, decides what may become action and participates in a shared world. It treats intelligence as a bounded process rather than a model floating outside its environment.

An ARCANA agent inside a shared worldSensed evidence enters a protected agent boundary. Inside, the agent combines possibilities with its current capacities. Only an authorised proposal crosses into the shared world; consequence returns as experience.senseintegratepossibilitycapacityshared worldagent boundarysensed evidenceauthorised actionconsequence returns
One agent inside a shared field. Evidence may enter its private process; only an explicitly authorised proposal becomes public action.

The shift

Memory, tools and multiple agents are usually added around a model. ARCANA starts one level earlier.

What counts as a thing? What survives change? Who is allowed to act? How does a private experience become shared evidence? The framework turns those questions into a deterministic graph-machine whose state transitions can be reproduced, inspected and challenged.

A common engineering stack

model → memory store → tools → multi-agent orchestration

Useful components, but their relationships, authority and continuity are often supplied by the surrounding application.

ARCANA's starting point

relation → boundary → permitted action → shared consequence → continuity

The components arise inside one typed process, and no stage silently acquires the authority of the next.

An agent episode

The agent does not move through a checklist. One relational state changes shape.

Sensing, considering, authorising and learning are different views of one evolving structure. Follow a single request as possibilities open, meet a boundary, become public consequence and change what can happen next.

One event five moments

One action unfolding over a changing relational fieldEvidence enters a bounded agent and changes its relational state. Several responses become possible, but only an authorised response crosses into the shared world. The witnessed consequence returns and changes the agent's next state.shared worldagent boundary
relational stateembodied boundaryshared consequence

Moment 01 of 05

A request changes the field.

A collaborator asks whether some private work can be shared. The request enters as evidence: it can affect the agent, but it does not yet carry permission or determine a response.

Sensing, considering, authorising and learning are different moments in one evolving structure—not separate tools passing a message along a pipeline.

One circuit, five movements

From a field of relationships to an agent that can continue.

Read downward as one cycle. Continuity is not the endpoint: what an agent carries changes the relations available on the next pass.

  1. Relation

    Relations come before isolated objects.

    ARCANA does not begin with permanent things that are connected afterwards. It rebuilds a field of relations from current state. Patterns that remain present, coherent and stable can crystallise into new things—and those things can enter later relations in turn.

  2. Boundary

    An agent is defined by what may cross its boundary.

    An agent is not a disembodied decision-maker. It has a private interior, a boundary that senses and protects, capacities that constrain what is possible, and a process for integrating competing pressures before anything becomes action.

  3. Action

    A proposal is not permission.

    The space of thought can remain wide and exploratory. Action is narrower because it creates consequence. An outward effect must pass the body, available resources, care, law or policy, and deliberate authorisation. Refusal is a valid result, not a system failure.

  4. Shared world

    Private experience becomes shared consequence.

    An agent's interior stays private. A world learns through voluntary disclosure and the witnessed consequences of action. It can retain evidence, coordinate care, govern shared resources and remember unresolved harm—without reading minds or turning a record into truth.

  5. Continuity

    Memory is carried change, not merely retrieved content.

    The native memory of an agent is the path-dependent change it carries into the next moment. Storage can preserve declared state, but learning, recovery, identity, public history and current authority remain different questions. A digest can verify what was kept; it cannot recreate what was lost.

The distinctions do the work

Similar words often hide different relationships.

Thought action
Reasoning can recommend a possibility. It cannot make that possibility cross into consequence.
Memory storage
Experience changes the agent. A database only determines which declared part can survive interruption.
Record truth
Public evidence prevents forgetting. It does not become a finding, a cause or a command by being stored.
Replay correctness
Exact reproduction establishes what happened. A wrong rule can also be reproduced exactly.

What it is

A runnable model of internal organisation.

ARCANA is a deterministic symbolic graph-machine. It models how an individual or collective is organised, how pressure circulates, how possibilities become legible and how experience changes what can happen next. Its canonical path is exact and replayable; faster Rust and CUDA paths can be checked against that reference.

What it refuses

A simulation is not the thing it describes.

Describing a bounded agent does not claim consciousness. Internal coherence does not predict external physical measurements. A replay is not real-world confirmation, and a readout is not permission to act. These limits travel with the results instead of being added later as disclaimers.