QUIVER: A Formal Framework for Quantifying Perturbation Propagation and Bifurcation in Compound AI Systems

In three linesQUIVER is a formal framework for measuring perturbation propagation in graph-structured LLM pipelines. It defines sensitivity matrices, trajectory divergence, bifurcation thresholds, and distribution faithfulness. Validated on 8,200+ instrumented traces across three distinct architectures.

Summary generated by Claude — human-verified