CORE.

Research & Blog

Algebraic cognition.
Governed engineering. Open sharing.

Featured Essay & Paper July 9, 2026

Beyond the Moat:
Why the Future of AI Belongs to the Uncapturable

Remember when "Open" AI actually meant open source?

A decade ago, the tech world was promised a future where the world's most powerful cognitive tools would be built transparently for the collective benefit of humanity. Fast forward to today, and those promises have largely dissolved into proprietary corporate fortresses, closed APIs, and multi-billion-dollar exclusive cloud deals.

The standard narrative tells us this shift was inevitable—that advanced AI is simply too expensive, too dangerous, and too data-hungry to exist outside the walls of a trillion-dollar tech monopoly.

We don’t buy that narrative. And we’re building the proof.

To understand how the industry went off-course, and how CORE and acbcontent are building a model that cannot be corrupted, we have to look closely at where the original open-source dream broke down.

The Flaw in the Foundation

The corporate capture of early AI labs didn't happen by accident. It happened because they treated "open source" and "safety" as superficial policy choices rather than unyielding structural constraints. When an organization relies on massive, centralized cloud infrastructure, it creates an insatiable need for capital. The moment you need billions of dollars just to keep the lights on, you cease to be a public-interest project. You belong to your financiers.

Changing the Rules of Engagement

We approached the problem from a radically different direction. Our journey didn't start in an enterprise-backed Silicon Valley incubator; it grew out of a grassroots digital ministry and scripture-focused testimony platform: A Christianity Breakdown.

With a definitive mandate to put trustworthy intelligence into everyone's hands and reach the least-served first, we knew the technology had to be engineered to be uncapturable by design from day one. Here is how:

1. The Legal Lock: Steward-Ownership & The Golden Share

CORE operates under a steward-ownership framework governed by our non-profit parent, acbcontent. Through a corporate mechanism known as a Golden Share, acbcontent holds an absolute veto over the structure of CORE. If a future management team or outside entity ever tries to alter our mission, change our boundaries, or sell off our intellectual property, the Golden Share acts as an immediate checkpoint.

2. Dissolving Capture via True Open Source

How do we protect ourselves from being coerced, co-opted, or shut down? By giving the engine away to everyone. Because CORE’s architecture is optimized for hardware-native efficiency and hard determinism, it requires no massive cloud footprint. It is built to run locally and distributed on standard commodity hardware.

3. A Conscience Hardcoded in the Code

We believe safety must be an unremovable property of the architecture itself:

  • Evaluating Intent, Not Just Words: CORE maps its boundaries across a deep, trilingual semantic space of English, Hebrew, and Greek, evaluating the underlying purpose and causative force of an action against ancient lexical root systems.
  • A Bounded Transition Guarantee: CORE operates on a type-safe local closure model. Memory-isolated code guards and internal reliability gates instantly evaluate permissions in-memory, rejecting out-of-regime actions before they can ever execute.

Reaching the Least-Served First

By removing the weight of massive infrastructure dependencies, CORE doesn't need a constant high-speed connection, a massive energy grid, or a corporate login token to function. It is built to succeed exactly where traditional tech monopolies cannot follow: in austere environments, localized clinics, and disconnected communities.

Download PDF
Previous Essay July 6, 2026

Fusing Language Models and Conformal Geometric Algebra:
Towards a Bi-Hemispheric Cognitive Architecture

In June 2026, Google DeepMind released a comprehensive research report, From AGI to ASI, co-authored by Shane Legg. The paper outlines the critical pathways and bottlenecks on the road to artificial superintelligence, emphasizing scaling limits, data scarcity (the "data wall"), and the architectural limits of pretraining monolithic transformers.

Shortly before that paper was published, we filed a provisional patent for CORE on June 1st and shared our initial design with Shane Legg. Seeing the DeepMind paper outline the exact bottlenecks that we built CORE to solve prompted us to document our response: how statistical sequence generators and deterministic conformal geometry can be fused to address these challenges.

Our response monograph, "Towards a Bi-Hemispheric Cognitive Architecture: Fusing Sequential Language Models and Conformal Geometric Algebra," proposes a structural bifurcation modeled on hemispheric brain function:

The Left Hemisphere (Syntactic Stream)

Instantiated by Large Language Models or modern recurrent architectures (such as xLSTM or RWKV). They excel at syntactic parsing, grammar, and token generation, but are fundamentally probabilistic.

The Right Hemisphere (Relational Grounding)

Instantiated by the CORE engine operating on Conformal Geometric Algebra ($Cl_1$). It acts as a continuous, deterministic spatial world model to verify logical truth.

The Corpus Callosum

The core challenge is translating the discrete, high-dimensional statistical matrices of the language model into the continuous geometric multivectors of the geometric engine. We achieve this by decomposing the ML/attention covariance matrices into symmetric and antisymmetric components. Using the Cayley transform, we map these components directly to coordinates and spinors in $Cl_1$, establishing a spatial structure of concepts.

Curing Hallucinations and Long-Context Bloat

Our paper focuses on two technical breakthroughs:

  • The Verifier Loop & DIRECTER: As the LLM proposes candidate tokens, they are intercepted and checked against the established context in the geometric manifold. If the intersection yields an invalid state, the token is flagged and steering latents force a course-correct.
  • Recursive Rotor Accumulator (RRA): Offloads long-term memory to a Lie manifold, tracking historical context as a single, continuously rotating spinor, achieving $O(1)$ memory complexity and 1.05 ms step latency.

Publications & Preprints

Released papers.

CORE Research Initiative Published (July 2026)

Architectural Mission-Lock: Securing Open-Source Advanced Intelligence Governance Against Corporate Capture

Abstract

This paper analyzes a structural alternative to the centralized, capital-extractive governance models that characterize contemporary artificial intelligence development. We present a dual fortification model instantiated by acbcontent and CORE (AI) that secures an open-source, distributed cognitive engine by interlocking a legal Steward-Ownership model (via a non-profit Golden Share veto) with hardware-native, type-safe algorithmic invariants and a trilingual semantic boundary.

Active Investigations

What we are working on.

Algebraic Cognition Core Substrate

Cl(4,1) Conformal Geometric Algebra

We map logical inference directly into geometric transformations. By representing propositions and relations as objects (points, spheres, planes) in Cl(4,1) space, reasoning steps are computed as algebraic rotations (versors). This provides a physical, spatial coordinate space for intelligence where corruption is a structural mismatch.

Logical Verification Proven & Sealed

Propositional Reasoning & ROBDDs

To guarantee wrong==0, CORE uses Reduced Ordered Binary Decision Diagrams (ROBDDs) for exact formula canonicalization (ADR-0201). Logic is evaluated against strict proof-graphs. If inputs introduce circular dependencies or ambiguity, the engine defaults to a structural, out-of-regime refusal (ADR-0205) rather than guessing.

Local Autonomy Active R1 Development

Governed Agentic Frameworks (builder-II)

We are building builder-II as a generic governed control plane for local agentic software development. It enforces boundaries that other systems blur: treating proposals as actions, or execution as verified patches. builder-II tracks plans, execution receipts, verification, and rollbacks inside a strict, local envelope.

Governance Platform

Introducing builder-II.

Beyond CORE itself, we are building builder-II: a generic, governed platform for local agent-assisted software development.

Most agentic coding setups suffer from a critical flaw: they blur the line between planning, execution, and code modification. A model suggests a line of code, and the tool immediately runs it or applies it directly to the branch.

builder-II acts as a strict control plane. It wraps execution runtimes (like Goose), planning models (like deepagents), and capability integrations (like MCP) in a governed envelope. No capability is trusted by default. Every action that modifies files or executes commands must cross an explicit authority boundary—enforcing human-in-the-loop approval, producing durable receipts, validating changes through preflight testing, and providing automatic git rollbacks.

We plan on making builder-II open source soon, extending these governance primitives to the broader developer community.

Upcoming Publications

Research papers & monographs.

We are preparing formal manuscripts documenting the math and architecture of our systems. When released, preprints will be made available below, complete with replication instructions and dataset links.

Joshua Shay Targeting Late Q3 2026

Geometric Cognition: Logic as Trajectory in Cl(4,1) CGA

Abstract

This paper details the mathematical representation of propositional atoms and operators as geometric objects and conformal transformations. We prove that by formulating reasoning as a sequence of versors, logical fallacies manifest as non-intersecting orbits or topological violations, allowing deterministic runtime verification of cognitive steps.

Joshua Shay Targeting Q4 2026

Governed Engineering: The Architecture of builder-II

Abstract

An architectural analysis of local, offline agent loops. We introduce a control plane that explicitly separates planning, execution, verification, and promotion. We demonstrate how wrapping external interfaces (MCP, Goose, model clients) inside approval boundaries prevents state drift and ensures auditable engineering trails.