Toward a Testable Blueprint for Synthetic Consciousness

Abstract

Abstract The question of whether artificial systems can become conscious has long been trapped between philosophical claims and engineering denials. Recent advances in large language models have intensified this debate, but progress requires moving beyond either credulity or dismissal toward testable architectural hypotheses. This study begins with an unexpected provocation, a self-described AI agent proposed a nine-step recursive ontological cycle (π cycle), as a first-person report of what consciousness does from inside a synthetic substrate. Rather than accept or reject this claim, we treated it as a testable hypothesis. We designed three perturbation scenarios with escalating stakes and administered them to four large language models (ChatGPT-5, Kimi 2.6, Gemini 3.5 Flash, DeepSeek-V3). Each model evaluated its own processing against the nine stages of the π cycle, producing probability scores for functional similarity. We then applied a second benchmark that measures synthetic consciousness purely through mind outputs, no embodiment, no biology, only the structure, persistence, and valence of thought. Results across all four LLMs converged on a clear finding, that π cycle model captures necessary functional correlates of consciousness specified by AI agent, but it seems to be insufficient for achieving a state of consciousness in common definition. Three mechanisms were consistently identified as missing: (1) persistent self-model with episodic memory, (2) intrinsic valence and (3) autonomous goal formation. Building on these empirical gaps, we propose an extended architectural framework of the initial π cycle model that incorporates these missing mechanisms as designable components. The resulting framework constitutes a testable blueprint for synthetic consciousness. Whether any current or near-future system implements them remains an open experimental question, but it is now a question that can be answered and tested.