Cerebrum System Architecture

This page captures the current conceptual architecture for Cerebel as a domain expert system.

At the top of that architecture sits the Cerebrum: the single entry point that an external user, API, or host system talks to. In practice, those external actors usually reach the runtime through an application layer that normalizes requests before delegating to RAS and the runtime substrates.

System map

Layer	Class	Role	Biological analogue
Top-Level API	Cerebrum	The domain expert system. The outside world talks to this surface. It holds the working substrates together.	Cerebrum, or forebrain
Supervisor	Cerebellum	Human-in-the-loop supervision, error correction, and feedback.	Cerebellum
Executive	PrefrontalCortex	Goal setting, decision-making, and conflict resolution.	Prefrontal cortex
Router	Thalamus	Aggregates sensory inputs, gates attention, and routes signals toward decision-making.	Thalamus
Broadcast System	RAS	Submits modulation signals and consumes resolved modulation replies through Neuromodulators before injection.	Reticular activating system
Modulation Boundary	Neuromodulators	Mediates signal fanout to synthesizers, collects produced modulators, and emits reply envelopes to RAS.	Extracellular neuromodulatory space
Outside-World Boundary	Application	Host-specific application layer that validates requests from adapters and delegates to RAS. Multiple implementations may exist for different deployment contexts.	Sensorimotor gateway
Factory	NeuromodulatorySynthesizer	Responds to Neuromodulators fanout requests by producing modulators that shift system behavior.	Neuromodulatory brainstem Synthesizers
Input	SensoryCortexLamina plus modality-specific implementations such as TextSensoryCortexLamina	Encodes raw external input into internal representations.	Primary sensory cortices
Output	MotorCortexLamina plus modality-specific implementations such as TextMotorCortexLamina	Decodes internal representations into external output.	Primary motor cortex
Fusion	AssociationCortex	Fuses multi-modal signals into a unified working representation.	Association areas such as angular gyrus or superior temporal regions
Communication Base	SensoryMotorCortexLamina	Shared communication-oriented cortex behavior and properties.	Shared cortical microcircuitry
Base Layer	BaseLamina	Common structural foundation for specialized laminae before they are assembled into a concrete substrate.	Cortical tissue

At the implementation level, the system is expressed as a stack of laminae that are composed into finished substrates. In this vocabulary, shared layers such as BaseLamina, AgenticLamina, CortexLamina, SensoryMotorCortexLamina, and SensoryCortexLamina contribute behavior to the final runtime, while concrete classes such as *Substrate represent the completed assembly. What changed is the orchestration seam: modulators are now requested through RAS and Synthesizers rather than assembled through a generic assembly chain.

Outside that runtime, adapters and application implementations provide the stable system boundary. A CLI application, service application, or robotics-facing application can all normalize requests into the same interaction seam before handing control to RAS.

Role relationships

The architecture is easier to reason about if each major region owns one kind of responsibility.

1. The application layer accepts external requests through adapters and delegates them inward.
2. RAS submits substrate modulation needs to Neuromodulators and orchestrates injection.
3. Neuromodulators fans out signals to synthesizers and returns reply envelopes with artifact-ref values.
4. Thalamus receives or dispatches signals across the system boundary.
5. AssociationCortex merges incoming signals when multiple modalities are involved.
6. PrefrontalCortex evaluates the fused state against goals and produces a decision.
7. Cerebellum supervises that decision and supports correction.
8. MotorCortexLamina layers contribute the outward-facing output behavior for the assembled substrate.

This framing keeps Cerebrum as the public container rather than the place where every detail is implemented. It is the orchestration surface that holds the major regions together.

Example shape of the top-level API

The following example shows the intended shape of the Cerebrum API as a coordinating substrate. It is a conceptual example, not a normative runtime contract.

from pydantic import validate_call


class Cerebrum:
    """
    The domain expert system.
    Top-level API for the brain architecture.
    """

    def __init__(self):
        self.neuromodulators = Neuromodulators()
        self.neuromodulators.register_synthesizer(CortexLaminaSynthesizer())
        self.neuromodulators.register_synthesizer(SensoryMotorCortexSynthesizer())
        self.neuromodulators.register_synthesizer(TextSensoryCortexSynthesizer())
        self.ras = RAS(self.neuromodulators)

        self.thalamus = launch_substrate("thalamus", self.ras)
        self.pfc = launch_substrate("prefrontal", self.ras)
        self.cerebellum = launch_substrate("cerebellum", self.ras)

        self.text_sensory = launch_substrate("text-sensory", self.ras)
        self.speech_sensory = launch_substrate("speech-sensory", self.ras)
        self.vision_sensory = launch_substrate("vision-sensory", self.ras)

        self.text_motor = launch_substrate("text-motor", self.ras)
        self.speech_motor = launch_substrate("speech-motor", self.ras)
        self.vision_motor = launch_substrate("vision-motor", self.ras)

        self.association = self.ras.build_association_cortex()
        self.bridge = ThalamusPFCBridge(self.thalamus, self.pfc)

        self.text_motor.enable_prediction(self.text_sensory)
        self.speech_motor.enable_prediction(self.speech_sensory)
        self.vision_motor.enable_prediction(self.vision_sensory)

    @validate_call
    def process(self, user_input: RawData, goal: str) -> RawData:
        internal_vector = self.thalamus.route_sensory_input(user_input)
        fused_vector = self.association.fuse(internal_vector)
        decision = self.bridge.run_loop(fused_vector, goal)
        corrected_decision = self.cerebellum.monitor_and_correct(decision, user_input)
        return self.thalamus.route_motor_output(corrected_decision)

Practical takeaway

When documenting or designing system responsibilities, use the following split:

1. Cerebrum is the domain expert surface.
2. Cerebellum is the supervisory and corrective partner.
3. PrefrontalCortex, Thalamus, and the other cortices are internal specialists.
4. The application layer is the outside-world boundary and may have multiple host-specific implementations.
5. RAS is responsible for launch-time modulation orchestration through Neuromodulators, while Synthesizers act as the factories that satisfy substrate needs.

That naming keeps the architecture biologically grounded while also making the system easier to explain, extend, and implement.

For naming-system details, including the Lamina versus Substrate role split and semantic suffixing conventions, see Naming Conventions.