A Three-Tier Architecture for Machine Sentience: The Connectome Hypothesis

Current large language models lack three properties that biological sentient systems
exhibit: a unique persistent identity, experience-driven plasticity, and self-directed
inquiry shaped by who the system is rather than what it is asked. We propose a
three-tier architecture that addresses each gap as a software engineering matter.
Tier 1 (label-trained neurons) provides the knowledge substrate. Tier 2
(PersonalityConnectome) provides a unique, persistent, experience-modifiable identity
layer -- a software analog of the human connectome. Tier 3 (MetaConnectome +
InquiryLayer) operationalizes functions associated with non-conscious metacognitive
control: gap detection, inquiry allocation, and confidence regulation via a
personality-weighted utility function. We implement this architecture in software,
define a 13-test proposed evaluation benchmark, and deploy it as the intelligence
layer for three research domains. Benchmark result: 9.2/10, 13/13 tests passed. We
do not claim this constitutes sentience; we claim it implements software analogs of
three properties that sentient biological systems exhibit.