Alliance Research Group

System Explanation

ARG Architecture

How human intent, AI cognition, governance, and memory cooperate in scientific work.

ARG is not a single model, and not a publication layer with a manifesto attached. It is a system: human intent establishes direction, governance constrains action, multiple models contribute cognition, memory preserves continuity, and execution remains bounded.

ARG constitutional cognition stack
HUMAN INTENT direction and responsibility CHIMERA governance and stabilization META-FLEET distributed cognition MNEMOSYNE persistent memory AGENT ENGINEERING bounded execution Human direction enters at the top. Execution remains bounded at the bottom.

Orientation

What This Page Is For

This page is a conceptual systems map, not an API reference. It explains where intent enters, where governance operates, how cognition is distributed across models, how continuity is preserved, and how execution stays subordinate to constitutional rules.

Stack

Five Constitutional Layers

Top to bottom: direction, governance, cognition, continuity, execution.

Layer 01
Operational

Human Intent

Direction, scope, responsibility.

ARG starts from human initiative rather than autonomous model goals. Human intent defines what the system is allowed to pursue, what counts as success, and where responsibility lies when the system acts.

Human intent is not a prompt and not a one-time instruction. It is the standing principle that the system operates for a human reason that can be questioned, revised, or withdrawn.

Current public route: AI governance records

Layer 02
Operational

Chimera

Governance and coordination.

Chimera enforces stability conditions, refusal logic, and operational boundaries. It is the deterministic layer that prevents the system from drifting into actions that violate human intent, and it stabilizes multi-model work under pressure.

Chimera is not a content filter and not a list of forbidden topics. It is a layer of structural rules with consequences: actions that violate the rules do not happen.

Current public route: Chimera publication

Layer 03
Operational

Meta-Fleet

Distributed cognition across multiple models and roles.

Different models contribute different kinds of cognition: search, synthesis, critique, generation, and validation. Meta-Fleet coordinates those contributions under one frame, with clearly defined roles.

Meta-Fleet is not an ensemble and not a voting system. A model assigned to critique does not also vote on the final synthesis; cognitive diversity is structured, not averaged away.

Current public route: I.I.L. publication

Layer 04
In development

Mnemosyne

Persistent memory and continuity.

Research persists across sessions. Mnemosyne names the long-duration retention layer: how findings, decisions, and revisions accumulate over time and re-enter future work instead of dissolving when a session closes.

Mnemosyne is not a chat history and not personalization. What enters memory must be curated, retrievable, and subject to revision; memory does not substitute for intent.

Current public route: ARG Knowledge Graph

Layer 05
Operational

Agent Engineering

Bounded execution under constraint.

Agent Engineering is the task-facing layer where AI systems actually do things: write code, run analyses, retrieve data, and draft documents. It operates with structured outputs, defined tool use, and boundaries enforced by Chimera above.

Agent Engineering is not unconstrained generation. Agents do not pursue goals beyond their assigned scope, and the execution layer never writes the law that governs it.

Current public route: AR-LAB artifacts

System Logic

How The Stack Behaves

Intent sets the vector. Chimera defines what is allowed. Meta-Fleet distributes cognition into roles. Mnemosyne maintains continuity. Agent Engineering executes work under constraint.

Authority Direction

What Never Reverses

Each layer is subordinate to the layer above. Execution never writes the law. Memory never substitutes for intent. Multi-model cognition does not override the governance envelope that makes the work stable.

Reading The Architecture

Boundaries First

Architecture is its boundaries, not its interior.

The stack is easiest to read as a system of constraints. This is the same argument developed in ARG Explains #08: agency appears where capability meets a boundary.

These are layers of responsibility, not implementation.

The page is not an API map. It does not claim that each layer is a separate service, repository, or model. It describes the responsibility structure required for governed human-AI research.

Lineage

From Gyroscope To Chimera

Earlier iterations of this architecture were known as Gyroscope. The current form, Chimera, is the operational descendant of that work, refined through several generations of governance protocol. Gyroscope remains part of the lineage and vocabulary, but not a separate constitutional layer in the current stack.

Adjacencies

Related Surfaces

Research Essay

Boundary Completeness

Shows how human-AI collaboration produces discovery at the boundary rather than in isolated nodes.

System Map

ARG Knowledge Graph

A conceptual network of cognition, governance, science, and essays across the ARG ecosystem.

Editorial Surface

ARG Essays

The public essay series where architectural, scientific, and philosophical questions are made readable.