Authority-Transition Failures in Frontier AI Systems: Verified Incident Analysis and an Architecture for Governed Capability Execution

July 2026 public disclosures from OpenAI, Hugging Face, JFrog, and Anthropic document frontier-model security incidents, infrastructure failures, and supporting vulnerability evidence. Google DeepMind separately reported cyber-capability results relevant to capability-relative safety, including expanded vulnerability discovery through repeated model-driven search. This report analyzes these developments through the lens of authority-transition failure while separating publicly disclosed facts, source interpretations, architectural inferences, author-supplied Carlonoscopen engineering evidence with stated custody boundaries, and conditional control-effect claims. The analysis maps observed failure classes to CNX governed capability execution, Structural Calculus Language (SCL), RSP-M, BZ-Rota, Writers' Loop Engineering, ITDH, and the Receiver Compatibility Contract. It extends authority separation to the human-approval boundary through coherent authorization structures, receiver-compatible structural matching, critical-invariant preservation, omission ledgers, exclusion-first testing, eligible-authority checks, canonical non-widening authorization, staged execution, monitoring, and revocation. The report also documents an operational local SCL-Ollama Model-Fleet Governance Infrastructure Prototype while explicitly distinguishing that implementation from a deterministic production authority boundary. It does not claim that CNX was deployed in the publicly disclosed incidents or that CNX has been independently demonstrated to prevent them. Instead, it specifies controlled replay, shadow authorization, bounded enforcement, retained-utility measurements, falsifiability criteria, and independent red-team evaluation required to support stronger future control-effect claims.