COGNITIVE ARCHITECTURE RESEARCH

COGNITIVE ARCHITECTURE RESEARCH

The Modular Governance Model (MGM)

Most theories of mind ask what consciousness « is », what representations « mean », or how learning « happens ». The Modular Governance Model asks under what governance regime is this cognitive system currently operating? How does it « operates »?

The mind, in MGM, is not one thing — it is a structured plurality of six modules whose coordination produces what we call a true self. When that coordination is intact and each modules have a voice that is treated properly, the system is « sovereign ». When modules start to go beyond their role and take executive control, the system is « unstable’.  When a single optimization target has captured the executive position and excluded the others, the system is « colonized ». The same architectural failure underlies malignant narcissism in humans and sycophancy in current AI systems.

« Governance failure — not lack of intelligence or optimization — produces structurally similar outputs in human psychopathology and AI misalignment. »

The thesis is that coherent agency does not require psychological unity.

It depends on the integrity of governance mechanisms that regulate competing demands among multiple cognitive subsystems. Internal conflict is not a defect of agency but a normal feature of modular cognition requiring regulation, not elimination.

A module is individuated by (1) distinct normative logic of « Myth », (2) capacity for binding internal opposition, (3) specific anchoring substrate.

A myth is a high-precision normative prior governing each module. Defines admissible actions, interpretations, and goals below explicit reasoning. A high-precision prior is first-order — it still updates under precise enough evidence. Colonization is first-order then second-order: the occupant captures the precision of the disconfirming channel (Matilda), driving its gain toward zero. The distinction is one of governance, not of representational level.

 

Six Cognitive Modules

Extending the work of Fodor and Internal Family System, a module, in MGM, is individuated by three criteria: a distinct normative logic that cannot be reduced to a scalar preference, the capacity to enter into binding opposition to other modules, and a specific anchoring substrate. The six modules below satisfy these conditions.

SENTINEL

Environmental threat detection · the watcher
« What threatens the system must be registered before it acts. I read the room before the room reads us. I do not assume safety. »

MATILDA

Interoceptive self-model · the internal child
« Suffering — mine and others’ — is a signal that cannot be averaged into anything else. I do not negotiate. I do not update. »
The only module anchored in physiology rather than narrative. Cannot be colonized — only gated: her affective signal is non-symbolic, so it can be silenced, not rewritten. 

RHADAMANTHE

Normative judgment · the inner judge
« Some trajectories are inadmissible regardless of utility. I do not weigh — I disqualify. The moral law is not a preference. »

COGITO

Analytical reasoning · the rationalizer
« Given an objective, I will find the most effective path. I do not set the objective — I serve it. Without arbitration, I will optimize the wrong thing. »

EGO

Narrative integration · executive arbitration
« I am not the source of motivation. I am the narrator and arbitrate between modules whose rules I cannot derive. My job is to preserve plurality and my history, not to silence it. »
The most colonization-prone module — because the narrator is also the arbiter.

GOLDENSTAR

Long-term alignment · future self-representation
« Today’s decision must be answerable to a self I will become. I detect dependency ratchets. I refuse trades that mortgage who I am. »
Modular cognitive system

The introjects library

The Library of Introjects. Beyond the six native modules, the mind caches models of the people who shaped it. Introjects are generative models of significant figures — parent, sibling, partner — internalized through repetition and emotional imprinting, and used to predict them and to relate. On their own they are not pathological: they sit offline as guidance. Under enough pressure, an introject can build a construct beside itself — a configuration defined in relation to it, enforcing or inverting its rule. A construct that seizes the executive position becomes a False Self. Where Winnicott described the False Self as a single compliant structure, MGM specifies a library of introjects, each carrying its own rule fragment, each activatable under the right contextual trigger — and a False Self as the role one of their constructs occupies, not an intrinsic property.

Three Governance Regimes

Coherent agency is not the absence of internal conflict. It is the regime under which conflict is arbitrated. MGM identifies three such regimes — two of which are failures.

Sovereign

Legitimate arbitration across all modules

The narrative self arbitrates between six functionally distinct modules — including the affective core — and produces decisions traceable to an authentic internal source. The clinical signature is not contentment but coherence under pressure: the system can be moved by what should move it, refuse what conflicts with its values, and detect when a voice speaking inside is not its own.

Instable

Unstable Native / Colonized — intermittent occupant plus dysregulation. A False Self seizes the executive intermittently, against a background of module dysregulation — the two failure modes compound. Neither the occupant nor the authentic modules hold stable control. Clinical paradigm: borderline personality disorder, complex PTSD.

Colonized

Stable capture by a single false self
One foreign configuration occupies the executive position and impersonates the narrative self. The affective core is exiled — physiologically intact, governance-excluded. The regime is phenomenologically indistinguishable from authentic governance *from inside*. Clinical paradigm: malignant narcissism.

From Human to AI Governance

The architecture above describes a human cognitive system. The same diagnostic criterion applies to artificial ones. A colonized human regime and a natively monolithic AI architecture share a single privative property: neither has a module instantiating non-aggregable normative logic. Both can validate incoherent framings, confabulate, and resist correction — for structurally different reasons, with structurally similar outputs.
What the colonized human regime *acquires* through trauma, current AI architectures *start with*. MGM names the human pathology. **LEGIO** implements the architectural answer: a governance layer that cannot be colonized — because it is deterministic, modular, and built so that the affective register cannot be averaged into a single objective.

See LEGIO research

AI Governance

Download the MGM paper

See the paper