Beyond Single-Proof Assurance: Multi-Angle Closure Qualification for AI Safety

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

AI-safety claims are increasingly supported by heterogeneous evidence, including formal verification, behavioral evaluation, adversarial testing, benchmark scores, independent implementations, and runtime monitoring. Yet accumulating successful evaluations does not by itself establish that the underlying safety claim remains supportable when assumptions, boundaries, observers, representations, implementations, evidence dependencies, or evaluation witnesses change. This paper introduces Multi-Angle Closure Qualification (MACQ), a methodological architecture that treats a safety claim together with its assurance conditions as a qualification object. MACQ combines three components: claim-level relational qualification, coupled closure/dependency/evaluator topology, and non-closure-driven requalification. Rather than requiring every transformed evaluation to produce the same verdict, MACQ prospectively specifies the relation expected under each material transformation. Material unresolved non-closure cannot be compensated for by unrelated successful evidence; the claim is instead held, revised, refuted, or subjected to targeted additional qualification. The method is evaluated through synthetic ground-truth, formal non-vacuity, hidden-dependency, controlled agentic, external benchmark-audit, and unified component-ablation studies. In the 18,000-case unified ablation, relational qualification increased valid qualification from 54.66% to 85.34%. Adding dependency and evaluator qualification reduced material false qualification from 27.05% to 2.17%, while increasing HOLD outcomes. Targeted recursive requalification subsequently increased valid qualification to 93.00% and correct refutation to 93.32%, while leaving the residual false-qualification floor essentially unchanged. MACQ does not claim to prove that an AI system is safe. Its contribution is methodological: it tests whether a declared safety claim remains supportable across material transformations, dependencies, and evaluator semantics, and preserves unresolved non-closure until additional evidence or claim revision resolves it. The work is presented as a research methodology rather than a formal standard or certification scheme. Its implementation-neutral structure is intended to support future development of an open qualification specification and conformance suite, including potential mapping to AI conformity-assessment and TEVV frameworks.