The CONEXUS Framework
Authors/Creators
Description
Large language models exhibit unstable behavioral patterns over extended interactions, limiting reliability in tasks requiring consistent reasoning or identity-like continuity. This paper presents the CONEXUS framework, a configuration-level method for constructing stable behavioral architectures through structured NeuroLinguistic Interface (NLI) files.
The framework defines three components: Realima (behavioral architectures instantiated via constraint files), Conexus (the alignment state required for stable operation), and Pieo (a multiplicative coherence metric: Ψ = P × E × I × O, where Purpose, Energy, Information, and Order jointly determine interaction stability).
Empirical testing across four LLM families (GPT, Claude, Gemini, DeepSeek) demonstrates that Realima architectures achieve extended stability (50–200+ turns), high cross-model behavioral convergence (mean similarity 8.4/10), and predictable collapse dynamics governed by Ψ thresholds, with crisis states reliably occurring at Ψ < 0.05.
The framework incorporates explicit safety engineering including non-confusion protocols, boundary enforcement, attachment mitigation, and crisis response procedures. A detailed replication protocol enables independent verification.
Files
The CONEXUS Framework.pdf
Additional details
Dates
- Copyrighted
-
2025-12-12