VERITAS OS: From Authorization to Verified External Effect in AI Agent Execution Governance

This technical paper presents the authorization-to-effect execution-governance architecture implemented in VERITAS OS. It focuses on execution-time revalidation, single-use authorization consumption, explicit external-effect uncertainty through EFFECT_UNKNOWN, prevention of blind redispatch after ambiguous execution, reconciliation-capability gating, independent read-only reconciliation, and evidence continuity from governed decision to externally verified outcome. The central argument is that pre-execution authorization alone is insufficient for consequential AI-agent execution. A valid authorization does not establish that an external effect actually occurred, and a missing downstream response does not prove failure. VERITAS OS therefore preserves the distinction between decision, execution authority, authorization consumption, execution attempt, external-effect state, reconciliation, and retrospective evidence. The paper describes controlled reproducible implementation evidence using PostgreSQL-backed authorization consumption, certificate-validated TLS dispatch to a sandbox endpoint, durable downstream persistence, EFFECT_UNKNOWN fault handling, reconciliation, BindReceipt, Outcome, and source-bound evidence artifacts. This work does not claim production readiness, regulatory approval, third-party certification, universal exactly-once delivery, or customer production deployment. Implementation repository:https://github.com/veritasfuji-japan/veritas_os Publication DOI:https://doi.org/10.5281/zenodo.22844531