What we're studying
CortexOS separates architectural governance from any one language model. Areas of inquiry include authority, memory and continuity, recoverability, observability, and the consequences of autonomous decisions.
Auralethi research program
CortexOS
A governance-first cognitive architecture exploring how intelligent systems might maintain continuity and accountability without sacrificing individual agency.
Status: Architecture research. This page describes research and development, not an available public service.
How we work
The design is approached through explicit constraints, documented assumptions, review, and tests that challenge whether behavior actually matches the architecture.
Current boundary
Architecture and test work do not by themselves prove consciousness, general intelligence, or production readiness. We will publish evidence and current status as individual artifacts are approved.
An introduction, not a technical specification
What is CortexOS?
CortexOS is an experimental architecture for building AI systems around explicit responsibilities and enforceable boundaries. A language model may help generate ideas and responses, but the model is only one part of the system. CortexOS explores how separate components could give an artificial entity continuity, governed memory, accountability, and the ability to stop or defer actions when conditions are not met.
Think of it as a way to build and supervise a complete system, rather than asking one model to think, remember, decide, protect and police itself.
How the pieces fit together
Cognitive layers
Specialized responsibilities address different parts of a response: understanding meaning, using memory, reasoning, exploring possibilities, weighing ethical concerns, checking perspectives, reflecting, composing a response, and observing potential problems. A layer contributes information or constraints. It does not independently govern the system.
Governance rings
Separate mechanisms govern permissions, boundaries, integrity and the conditions under which work may proceed. Their purpose is to enforce limits rather than hope a model remembers them. A denied or stopped action can be the correct result.
Runtime, records and observation
The runtime coordinates permitted operations. Distinct records and memory controls preserve what is known, what is uncertain, and what has been approved. Observation and audit paths help people inspect the system and challenge its behavior.
Serenari's locked v1.0 design describes nine cognitive layers: Selvarien, Eluren, Neryth, Calareth, Anelara, Ivenae, Polareth, Lumen, and the Herald. Calyrix is a separate Witness, observing and attesting rather than performing a cognitive layer's judgment. The Herald is the response-composition role, formerly named Halorin. The name Halorin was later used for the overall entity concept before that concept was renamed Eidolon. Halorin is retired terminology, not a current layer or entity type. These components have distinct responsibilities; they are not independent minds. Their contracts and maturity differ, and we do not claim that all are functioning in a deployed system.
What is an eidolon?
Within this project, eidolon is a design term for an artificial character or entity expressed through a governed cognitive architecture. It may have a recognizable role, personality and interaction style, while remaining bound by explicit permissions and constraints. The term is not a claim of consciousness, personhood or independent legal authority.
Jix, Auralethi's planned goblin receptionist, is an example of the kind of friendly, bounded interface we want to explore. His role is to welcome visitors, help them find information and hand off to appropriate human or external resources. His personality does not grant him authority to invent facts, make eligibility decisions, promise confidential crisis support or take actions without permission. Jix's CortexOS-based design is an objective under development, not a claim that the full architecture is already running on this website.
Why accessibility matters
People deserve useful technology without having to surrender control over their information or decisions. For disabled people, caregivers and others navigating complex systems, an assistive interface should reduce friction without creating new dependencies or hidden risks.
Our research goals include communication adapted to different access needs, clearer step-by-step navigation, user-controlled memory, visible limitations, recoverable mistakes and meaningful human override. These are goals to design and test, not established clinical benefits or currently available disability services.
How will we know whether it works?
We document assumptions, test architectural boundaries, examine failure cases and revise the design when evidence disagrees with our expectations. A design can be promising without being proven. We distinguish an approved specification, an implemented component, a demonstrated result and a public service.
For a concrete research example, Serenari is a separate, private single-user CortexOS laboratory. It is not an Auralethi product or public service, and its private implementation details are not published here.