Shared Semantics without Shared Authority: A Design Analysis of the Agent Capability Contract

Working paper, version 1.0 (4 November 2026). Not peer reviewed. This paper analyzes Agent Capability Contract (ACC), a capability-declaration design that shares selected agent-facing semantics while retaining final authorization in business systems. Using a fixed public specification, comparisons with adjacent mechanisms, six constructed cases, and a reproducible reference-checker defect, it distinguishes declarations, effective decisions, and evidence of outcomes. The original reference checker passes its 26 published vectors but misclassifies a valid combination of an OpenAPI GET operation and an explicit non-readonly declaration. A subsequent correction preserves the existing rule and expands risk-precedence coverage to 24 combinations within a 48-vector corpus. These observations establish a specific reference-implementation defect and its correction, not a production vulnerability or end-to-end security result. The contribution is a testable account of the handoffs surrounding capability declarations, together with explicit conditions under which a declaration layer may help, duplicate existing mechanisms, or remain insufficient. The analysis does not change ACC normative requirements. Author: Jingchuan Nie, independent researcher and ACC maintainer. Substantive AI assistance in drafting, translation, document analysis, and internal critical review is disclosed in the paper. This assistance is not independent peer review.