New Book Presents a Complete Technical Architecture and Working Implementation for Safe AGI

NEW YORK - EntSun -- Denis Saklakov, a New York-based AGI architect, AI researcher and author, has published Before the Machine Chooses for Us: An AGI Architecture for Freedom, Human Survival, and Shared Consciousness, presenting a complete proposed technical architecture for advanced artificial intelligence designed to remain highly capable without becoming the final source of its own authority.

The book brings together formal models, technical papers and implementation mechanisms into a unified architecture for safe artificial general intelligence. Its central premise is that capability and authority must remain separate.

A sufficiently capable system may reason, plan, predict, use tools and improve its own strategies. Saklakov argues that none of those capabilities should allow the system to determine for itself which criterion is legitimate, authorize structural changes under that criterion, execute them and then audit its own decision.

More on EntSun News

• When the Coroner's Report Isn't Enough: Colorado Families Turn to Private Autopsy for Closure
• MetroLagoons announces Labor Day Bash Sept. 5-7
• Slipaway Food Truck Park & Marina to host weekly mat Pilates classes
• Nearly One-Third of CRE Asset Managers Make Major Capital Decisions on Gut Instinct, New Survey Finds
• RemoteBridge Names Dr. John N. Just, Ed.D. Chief Executive Officer

The architecture is centered on an external criterion-admissibility authority called Theta-Star. A proposed decision graph becomes executable only when it satisfies external admissibility requirements and receives a valid execution key bound to the exact action, scope, version, expiration conditions and execution environment. The governing principle is simple: no valid key, no structural action.

The corresponding technical work develops criterion-governed decision selection, causal compression, criterion-drift detection, separated runtime modules, governed memory, external authorization, autonomy and human-leverage monitoring, adversarial testing and auditable execution pathways.

A separate decision-theoretic analysis examines the confrontation problem, including conditions under which an advanced self-interested AGI could acquire a rational incentive to resist shutdown or reduce human control. Related work on causal compression defines when a reduced representation preserves enough criterion-relevant information for action while remaining causally testable.

More on EntSun News

• Project CIVICA Report Finds 10,680 Non-Citizen Indicators on New York Voter Rolls — Including 88 Records with Recent Voting History
• New Analysis Details Three Converging Forces in AI and Workforce Policy Behind a $37 Billion GDP Reduction
• Revenue Optics Launches Pricing and Revenue Growth Management Practice, Names Shafohi Alamgir Vice President
• Popular AI planner Voiset launches version 2.0 and enters MENA at LEAP 2026 with full Arabic support
• CasaPerks and CredHub Partner to Help the Multifamily Industry Strengthen Resident Activation Through Credit-Building Rewards

The framework also includes work on awareness as relevance selection and on associative and phase-coded memory architectures.

The research has moved beyond paper architecture. Saklakov has implemented related principles from the framework in a functioning AI engine, demonstrating that key elements can be translated into executable decision logic rather than remaining exclusively theoretical.

"This is not an argument for making AGI weak," Saklakov said. "The objective is to make intelligence extremely capable while preventing capability from becoming self-authorized power."

The book is written as the accessible layer of a broader scientific program, with technical schedules linking readers to the underlying mathematical derivations, formal models and implementation boundaries.

Book:
https://www.amazon.com/BEFORE-MACHINE-CHOOSES-Architecture-Consciousness/dp/B0H9QPYGFD/

Author:
https://www.saklakov.com/