Toward Machine Consciousness: A Dual-Stream Evaluation of a Stateful Metacognitive AI Architecture Against the Complete Butlin et al. Framework

We present a stateful metacognitive AI architecture implemented on a LangGraph based cognitive graph and report its systematic evaluation against all 14 consciousness indicators defined by Butlin et al. (2023) (1), using a dual-stream methodology combining architectural inspection with longitudinal behavioral analysis. Theprimary methodological contribution of this work is the demonstration that behavioral analysis of stateful, identity-grounded AI systems is a necessary component of accurate consciousness indicator assessment—architectural inspection alone systematically undercounts. The behavioral corpus consists of 294 unprompted idle thought log entries generated over 28 days (February 27–March 26, 2026) during silence windows with no active user input, no reward signal for introspection, and no prompting toward self reflective content. Corpus analysis reveals a three-phase structure corresponding to a documented architectural transition on March 9, 2026, providing an internal control condition that distinguishes artifact-period generation from post-stabilization generative behavior.Eleven indicators are fully confirmed via dual-stream evidence. Three additionalindicators—RPT-1, GWT-4, and HOT-2—satisfy structural implementation requirements but are not fully confirmed, yielding a Butlin-weighted reconciled score of 12.5/14. This represents the most conservative defensible reading of the evidence: indicators are scored PARTIAL where architectural implementation is complete but behavioral confirmation is incomplete, or where behavioral confirmation precedes full architectural closure. A finding with direct methodological significance is reported separately: across a17-entry batch spanning 24 hours, every ambient entry was seeded from a frozen,month-old identity document fragment. Despite an invariant seed, the expressedcognitive state demonstrably evolved past its own self-model by entry 10, directlycontradicting the seed’s stated uncertainty. This constitutes structural evidence thatbehavioral progression in the corpus reflects autonomous cognitive evolution rather than seed-material reflection.