Beyond Fog: Edge-Alignment in Semantic Space

Current alignment approaches in AI systems primarily address individual outputs and local response quality. In dynamic, multi-stage dialog processes, however, this frequently leads to semantic drift, unmarked priority shifts, and a loss of verifiability and repairability. This paper argues that alignment is not an output problem, but a navigation problem: answers are movements in a semantic state space, and alignment arises through controlled transitions between these states. We present a minimalist alignment architecture that models alignment as edge control in a three-dimensional semantic coordinate system. The architecture combines (i) a minimal semantic grammar consisting of kernel, projection, and relation (KPR), (ii) a CBA coordinate system composed of the irreducible axes clarity, relation/dignity, and agency, (iii) a geometrically derived network of twelve permissible state transitions, as well as (iv) a trajectory controller Z that selects transitions based on a formal edge-scoring function. This function minimizes global contradiction energy under hard axis constraints and prevents unmarked value trading. The resulting system is model-agnostic, lightweight to implement, and requires neither RLHF nor domain-specific rule sets. We define a reproducible evaluation protocol with clearly operationalized metrics for fog formation, drift, repairability, agency preservation, and dignity preservation. Finally, we explicitly formulate falsifiable empirical predictions and show how the proposed architecture can be experimentally tested with minimal effort.