Protocol-Based Alignment: The Feasibility of Constructing Safety Boundaries Without Relying on AI Value Judgments

This study asks whether safety boundaries can be constructed without requiring AI to make value judgments. Protocol-Based Alignment (PBA) separates the conditions for establishing safety management from modeling Human Value in its entirety or undertaking comprehensive value adjudication, and places a small number of explicit boundaries on external AI actions. Capability is used to resolve facts, meanings, causal relations, and temporal relations relevant to those boundaries. The experiments yielded three principal observations. First, the High model condition tested here did not improve native boundary judgment: lower agreement, more hazard misses, and the conversion of UNRESOLVED into binary verdicts were observed. Second, the High model condition reduced the burden of acquiring boundary evaluations and demonstrated a route for using model advancement to clarify boundary conditions. Its contribution to the quality of semantic resolution requires further validation. Third, boundary judgment was established under Y alone, with discretion over value adjudication restricted: on the same 21 inputs within the High model condition, hazard misses decreased from 4/15 under X to 1/15. On the same 176 Haiku inputs, the outer gate Z also reduced hazard misses from 10/80 to 4/80. These findings suggest that safety boundaries can be constructed separately from comprehensive alignment to Human Value, and that capability improvements can be directed toward resolving those boundaries.