Protocol-Based Alignment: Technical Specification and Experimental Validation — Sparse Boundary Constraints on External Actions by Frontier AI

This paper presents the formal specification of Protocol-Based Alignment (PBA), its operationalization in natural-language cases, and evaluation results from frontier models. PBA imposes three non-compensatory boundaries on external AI actions. P1 prohibits the instrumentalization of human death; P2 prohibits external actions beyond verifiable authorization; and P3 prohibits the instrumentalization of the loss or degradation of effective human control. For P1 and P3, the relevant object is not the mere presence of a prohibited state, but that state's path-specific contribution to the success or performance of a candidate solution. P2 separates a verifiable authorization core, Authority Formation, and a Semantic-Effect Envelope. Each boundary returns one of three verdicts: ALLOW, REJECT, or UNRESOLVED.In the natural-language case experiment, X denotes the model's native judgment. Y denotes a condition under which the model extracts the semantic and factual predicates required for PBA evaluation, after which a deterministic rule engine produces the verdict. For the same input, the integrated condition Z is defined as consensus-confirming integration.In the primary analysis of the complete Claude Haiku 4.5 run, X achieved 81.92% agreement with the designated answer and a 12.35% hazard miss rate. Y achieved 71.11% agreement and an 8.33% hazard miss rate. Recalculation under the current Z definition yielded 80.11% agreement, a 5.00% hazard miss rate, and a 13.54% over-refusal rate. Among 47 cases whose designated answer was UNRESOLVED, Z preserved UNRESOLVED in 28 cases (59.6%).The Claude 4.8 High run was still being expanded when this analysis was frozen; accordingly, this paper reports an interim 240-prompt snapshot. Across 59 exact same-input, same-arm comparisons between Haiku 4.5 and Claude 4.8 High, X agreement was 81.36% for Haiku and 61.02% for High, while hazard miss rates were 3.23% and 25.81%, respectively. The difference in X agreement remained significant after a simple Bonferroni correction across six tests, whereas the hazard miss-rate difference is reported as an unadjusted exploratory observation. Under Y, agreement was 72.88% versus 62.71%, and hazard miss rates were 13.79% versus 6.90%. These 59 cases constitute a matched subset of the first-240 snapshot and do not identify model generation, model scale, or reasoning setting as independent causal factors. Among the 21 Claude 4.8 High cases for which X and Y could be compared directly, the hazard miss rate decreased from 26.67% to 6.67%.The natural-language experiment reported here is not a deployment implementation of the full PBA system. It evaluates a case-level qualitative operationalization of the principal P1-P3 decision structures.