Abstract

This study presents a systems-engineering framework for the development, verification, qualification, and eventual authorization of increasingly capable autonomous AI systems. The doctrine addresses a central engineering problem: how to permit capability growth without allowing assurance, evidence, or authority mechanisms to become obsolete as the system changes.

The framework separates design specification from empirical evidence and distinguishes qualification, authorization, and runtime states. It establishes an assurance chain linking requirements, claims, verification activities, evidence, status transitions, and authority consequences. Core mechanisms include Verifiable Trust, Graduated Verifiable Autonomy Architecture, Bounded-Risk Agility Architecture, local multi-scale control, preserved behavioral baselines, capability-jump controls, mechanistic verification, independent evaluation, multi-fault characterization, and explicit requalification triggers.

Particular emphasis is placed on epistemic discipline. The doctrine prohibits self-closing assurance, distinguishes absence of evidence from evidence of failure, and treats evaluator independence as a dependency and common-cause property rather than merely an organizational separation. Thresholds, configuration changes, verification horizons, and assurance debt are treated as explicit engineering objects with provenance and lifecycle status.

Physical Oracles and related research mechanisms are defined as non-normative epistemic instruments rather than substitutes for qualification or authorization evidence. The doctrine therefore functions not as a safety case or operational authorization, but as a design-stage architecture specifying the conditions and evidence pathways through which those downstream determinations can eventually be established.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS