Work

Agentic Competence Without Demonstrated Subjectivity Functional Relevance, Intervention, Memory, and the Limits of Synthetic Agency

Year: 2026
Type: preprint
Abstract: This paper presents a critical experimental and philosophical investigation of synthetic agency, originally motivated by the possibility that functional individuality might emerge prior to, or independently of, phenomenal consciousness. Drawing on 25 archived synthetic research pilots, the study examines how artificial agents acquire, preserve, revise, and sometimes misapply knowledge in changing environments. Seven principal experimental contrasts address active intervention, historical memory, costly information retention, learned state distinctions, temporal action composition, and the integration of competing needs. The experiments demonstrate that agents can obtain otherwise inaccessible information through their own actions and benefit from historically acquired representations. However, they also reveal significant limitations. Retained knowledge can become harmful, agents may construct distinctions from non-informative signals, and more elaborate architectures frequently fail to outperform simpler classical methods. Correctly recognising a change in functional relevance does not necessarily improve long-term regulation. These findings motivate a philosophical distinction between regulatory competence, historically conditioned functional relevance, autonomous normativity, and phenomenal subjectivity. The paper proposes relevance repair as a falsifiable research hypothesis: advanced agentic organisation may depend not merely on representing the world, but on maintaining and revising the conditions under which acquired knowledge remains meaningful for continued functioning, while preserving viable future alternatives. No claim of demonstrated synthetic consciousness, autonomous subjectivity, or a novel superior learning algorithm is made. Published as an open-science, unreviewed exploratory synthesis, the study is accompanied by a comprehensive evidence archive containing original source code, experimental data, protocols, negative controls, statistical analyses, and documentation supporting reproduction. Complete independent replication of all training runs has not been performed.
Language: English
Open Access status: green
OpenAlex ID: W7225905742
Cites:
Cited by:
Related to: 10