ACVAM-DA: An Open-Valence Metareflective Architecture for Artificial Autonomy and Consciousness Research

Résumé fourni par la source

This article introduces the Open-Valence Metareflective Cyclic Architecture with Autonomous Development (ACVAM-DA), a falsifiable conceptual framework for investigating artificial autonomy, metacognition, hierarchical self-modification, and theory-derived indicators of consciousness. The architecture refines a basic pleasure–displeasure regulatory loop into three distinct levels of valence: elemental positive and aversive signals, context-sensitive appraisal, and reflective endorsement. Positive and aversive components are represented independently as vector-valued states, preventing intense mixed-valence conditions from being obscured by scalar aggregation. During a bounded calibration phase, out-of-range elemental transducer states may trigger local parameter restoration. After calibration, however, mature preferences and appraisals are not automatically reset solely because their valence reaches extreme values. Instead, the architecture employs monitoring, temporary suspension, causal diagnosis, sandboxed modification, and reversible checkpoints. A minimal non-valential integrity kernel constrains external harm, observability, resource integrity, and reversibility while remaining conceptually separate from mutable preferences, goals, and short-term utility. ACVAM-DA further integrates recurrent global broadcast, generative world and self models, autobiographical memory, causal metacognition, endogenous goal formation, and four orders of plasticity. Two elementary information-loss results are established for scalar valence aggregation and unconditional reset mechanisms, an explicit acceptance rule for self-modifications is formalized, and a program of matched longitudinal ablation studies is specified. The proposed empirical framework distinguishes functional autonomy, metacognition, access-like global availability, phenomenal consciousness, and artificial life, and does not infer any of these properties from fluent self-report alone. A structured scoping search conducted through 29 August 2026 identified close precedents for each individual component of the proposal, including a particularly related recursive-feedback architecture, but no retrieved source instantiated the complete conjunction of the four operational claim bundles defined in the work. The calibration-to-open-valence transition emerged as the most discriminating feature. Accordingly, ACVAM-DA is presented as a potentially novel operational synthesis and a preregistrable experimental research program, rather than as evidence of absolute historical priority or as a demonstration that artificial consciousness can be engineered or detected with certainty.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.