We present EVE (Evidential Value-aligned Engine), a developmental cognitive architecture that integrates cross-domain task solving, grounded language acquisition, experiential preference formation, normative self-governance, algorithmic self-modification, and developmental self-reflection within a single tick-synchronous processing loop without any generative language model. EVE’s 15 subsystems share state through a unified cognitive loop, reusing core mechanisms across task solving, language, transfer, and governance. The paper reports nine headline results, including held-out transfer success, external benchmark performance across 3,123 tasks from five public suites, grounded language with zero confabulation, drive-mediated preference formation, runtime normative governance, recursive self-improvement, behavioral belief inference, learned representation-guided transfer, and endogenous goal generation. The contribution of the paper is the integration of these capabilities within one evaluated neuro-symbolic architecture together with quantitative evidence for each component and their interactions.
Matija Ludvig (Mon,) studied this question.