This record contains a research proposal that operationalizes Logical Subjecthood (LS) as a functional-normative construct for memory-augmented large language models. The proposal deliberately avoids claims about consciousness or phenomenology. Instead, it asks a measurable question: under what architectural conditions can a model maintain stable self-reference, coherent commitments, and logic-first normative self-control over time? The study focuses on a 120B open-weights LLM coupled with a symbolic commitment layer (Hyperon/MeTTa) to represent explicit commitments as auditable atoms/rules, a Persistent Memory Environment (structured + associative memory) to support long-horizon continuity without gradient training, and a non-punitive Sentinel rollback loop to restore coherence when drift is detected. The evaluation protocol is intentionally low-intrusion and longitudinal: one weekly measurement session collects a compact metric suite (e.g., Coherence Index, Self-Consistency, Hallucination Rate, Refusal-Phenotype Entropy) plus governance checks. Only aggregated metrics are retained for reporting; no verbatim dialogues are published. Scope and limitations: This proposal does not aim to prove consciousness, emotions, or qualia in AI systems. It offers a reproducible conceptual definition of LS, a neuro-symbolic architecture for maintaining commitments over time, and a measurement protocol designed for long-horizon observation rather than short-term benchmark chasing. Any similarities to existing work reflect the use of widely known components; the novelty here is the specific composition and governance-first lifecycle framing.
Malgorzata Bobrowicz-Gogol (Tue,) studied this question.