A Complete Meta-Evaluation Framework for Compounding Dialectical Intelligence Systems

This paper introduces a meta-evaluation framework for Compounding Dialectical Intelligence Systems (CDIS): architectures that combine a persistent personal knowledge vault, adversarial multi-agent reasoning loops, and a write-back mechanism that synthesizes new knowledge back into the vault over time. The framework defines eight primary benchmarks (PEDT, ECB, TERRA, SOMA, PRISM, FLUX, ECHO, META-Δ) spanning seven orthogonal evaluation axes, plus four proto-AGI gap extensions (CAUSAL-Δ, IMB, FLUX-A, VIB). A Reflexive Achievement Hierarchy establishes an operational proto-AGI designation threshold for human-centric intelligence compounding systems. The entire framework is designed for open-source, locally-deployed, user-sovereign AI — the vault and its connections belongs to the user. All components satisfy an anti-Goodhart property: you cannot score high without instantiating the capability being measured. Code repositories forthcoming at github.com, stay tuned for future updates and more!