The increasing complexity of modern AI systems exposes a significant assurance gap: safety evidence from practices like red-teaming and robustness testing remains fragmented, lacking a formal mechanism for composition and propagation throughout the development lifecycle. This prevents the construction of rigorous, dynamic safety cases essential for trustworthy AI. We introduce the Composable Assurance Framework (CAF), a novel engineering methodology that integrates safety assurance directly into MLOps workflows. At its core is the Formal Safety Assertion (FSA), a standardized, machine-readable structure that verifiably links safety properties—such as robustness scores or the absence of deceptive circuits—to specific AI artifacts. We then define a Composition Calculus, a set of formal rules governing how FSAs are propagated and aggregated as components are combined into a system. This approach transforms the development pipeline into an automated evidence-gathering engine, whose output is a dynamic Directed Acyclic Graph (DAG) of assertions that constitutes a living safety case. Through a prototype and a Retrieval-Augmented Generation (RAG) case study, we demonstrate how CAF automatically enforces a predefined safety policy, blocking non-compliant deployments.

Abstract The integration of precise static analysis into interactive development environments (IDEs) necessitates algorithms that can update analysis results incrementally within milliseconds. However, maintaining Bidirected Dyck-CFL reachability —the standard formalism for field-sensitive alias analysis—under dynamic graph mutations remains an open challenge. Standard batch algorithms exhibit prohibiting cubic complexity ( $$\mathcal {O}(N^3)$$ O ( N 3 ) ), while existing dynamic approaches often lack formal guarantees when handling non-monotonic edge deletions (the “ghost path” problem). In this paper, we present GC-DBDR , a novel incremental framework rooted in Abstract Interpretation. We reformulate the dynamic analysis problem not as graph patching, but as computing Differential Fixpoints over a lattice. By establishing a rigorous Galois Connection between execution traces and reachability relations, we derive update rules that are correct-by-construction . To resolve the asymmetry between monotonic insertions and non-monotonic deletions, we introduce a Counting-Augmented Abstract Domain supported by a Derivation Hypergraph . This structure operationalizes the Inverse Abstraction Principle , ensuring that reachability facts are retracted if and only if their supporting derivation trees are fully invalidated. We provide formal proofs demonstrating that GC-DBDR is sound and complete relative to batch analysis. Empirically, the algorithm achieves an optimal input-output complexity of $$\mathcal {O}(\varDelta )$$ O ( Δ ) , delivering sub-millisecond tail latencies on real-world benchmarks.