🔐 HASH PDF-a (ZA OPIS ILI README) SHA-256: b695b6c683d87d7ae1b262fb543f8e80fecd21f4961a8b6e8a09dc246a5b4c19 This publication discloses a formal architectural framework for pre-decision admissibility enforcement in AI systems, serving as a defensive prior-art disclosure. The work introduces a relational state-space formulation and a structural admissibility boundary (ΣE) that constrains which classes of system states are allowed to exist before decision or response generation. The contribution formally defines: • a system interaction state space (S) and a relationally inadmissible subspace (S_rel), • an admissibility predicate (ΣE) evaluated pre-decision, • a boundary projection mechanism (P) that preserves functional helpfulness while eliminating persistent relational dependency cues, • and a safety invariant ensuring inadmissible relational states never become reachable. Relational language is treated as a generator of state classes, not as prohibited content, distinguishing this approach from post-hoc moderation, alignment, or guardrail techniques. This publication intentionally omits implementation details, parameters, algorithms, and compute-core logic. It is released to establish prior art and prevent exclusive patent claims over the disclosed architectural concept. Scope note: This document defines architectural concepts and formal state-space constraints only. Concrete implementations, parameters, and compute-core logic remain proprietary and outside the scope of this disclosure. ⚖️ LICENCA (preporuka) ✅ Creative Commons Attribution 4.0 International (CC-BY-4.0)