Research guide
AI Engineering and AGI Engineering: Reliable Delivery, SCR and MRAC
In COG's research, AI engineering studies the conditions, mechanisms, methods and infrastructure for reliable delivery by intelligent systems in open, complex tasks with persistent constraints. Engineering AGI provides a long-term project-level autonomy objective connecting capability structure, process governance and verifiable results.
This guide introduces COG's research perspective. Consult the original papers for complete arguments, conditions and references.
An iterative audit and repair loop
How does AGI engineering relate to AI engineering?
AI engineering investigates how reliable delivery becomes possible; its practical value does not depend on first achieving AGI. Engineering AGI is an important long-term objective: enabling a complete system to sustain complex-project delivery with low human intervention under stated conditions, while extending its coverage. Here, AGI engineering refers to engineering work toward that objective. Building a tool or workflow alone does not establish that the objective has been met.
AI Engineering: A Disciplinary Program for the Reliable Delivery of Intelligent Tasks
What roles do KSC, SCR and MRAC play?
KSC supplies coordinates for knowledge, structure generation and constraint maintenance. SCR relates search space, coverage capability and delivery reliability, motivating search-space governance. MRAC supplies process observations of coverage state through audit-and-repair trajectories. The notation must remain distinct: S and C in KSC mean structure generation and constraint maintenance; in SCR, they mean search space and coverage capability.
AI Engineering: A Disciplinary Program for the Reliable Delivery of Intelligent Tasks
How does MRAC support reliable delivery?
MRAC observes changes in the number, severity and types of issues across repeated audits. Persistent non-convergence can signal that a task has not entered the system's stable operating range, supporting repair, rollback or decomposition. Convergence offers conditional process evidence that depends on continued audit effectiveness. It cannot rule out shared blind spots or replace independent acceptance checks.
How does MRAC Bench support AI engineering evaluation?
MRAC Bench applies MRAC to software engineering evaluation using tasks derived from real repositories and change requests. It examines convergence within budget, audit-and-repair trajectories and resource consumption. The paper proposes an evaluation method and infrastructure; its specification-audit pilot does not directly establish actual coverage or a general ranking of model capabilities.
Original papers and DOIs
AI Engineering: A Disciplinary Program for the Reliable Delivery of Intelligent Tasks
Large language models and their Agent systems can already generate code, plan tasks, invoke tools, and modify real engineering states, yet growth in capability has not been matched by corresponding improvements in the reliable delivery of c…
DOI: 10.5281/zenodo.21931156From the SCR Relationship to Search Space Governance: A Foundational Analytical Framework for Delivery Reliability in AI Software Engineering
Generative AI can already produce code and functionality at low cost, yet the reliable delivery of complex software still depends on continuously managing task constraints, component dependencies, change impacts, and validation feedback. Th…
DOI: 10.5281/zenodo.21917081Multi-Round Audit Convergence: Coverage-State Assessment and Recursive Control for Complex AI Engineering Tasks
Large language models and their agents can already handle some real-world software engineering tasks, yet a single successful execution is still insufficient to demonstrate that a specific result has reached a level at which it can be deliv…
DOI: 10.5281/zenodo.21959920The Theoretical Starting Point and Progressive Research Agenda of MRAC: From Non-Convergence Diagnosis to Engineering AGI
Previous work introduced Multi-Round Audit Convergence (MRAC), which evaluates whether a system can stably handle a current task by tracking how issues change across repeated “audit–remediation–re-audit” cycles on the same engineering objec…
DOI: 10.5281/zenodo.21917630MRAC Bench: A Software Engineering Evaluation Method and Infrastructure Based on Multi-Round Audit Convergence
Software engineering evaluation needs a continuing supply of tasks that reflect real development, distinguish system capabilities, and remain affordable to construct. This paper proposes MRAC Bench, an evaluation method and infrastructure b…
DOI: 10.5281/zenodo.22790766KSC: Knowledge, Structure, and Constraint — A Functional Framework for Human–AI Differences and Collaboration
Current AI systems can draw on broad bodies of knowledge and rapidly generate complex structures, yet their capabilities still vary substantially across tasks. Some problems are constrained mainly by the information available; some require …
DOI: 10.5281/zenodo.21991838An Engineering Definition of AGI: Complex-Project Closed-Loop Capability as an Assessment Criterion
Artificial general intelligence (AGI) has long lacked a stable and operational definition. Existing definitionsvariously emphasize human-like behavior, subjective mind, broad cognitive ability, human-levelperformance, economic value, or soc…
DOI: 10.5281/zenodo.22803294