Source record

Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2609.18820.

Pipeline notes

The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.

Paper defines Compositional Policy Violations (CPVs) in agentic AI workflows: individual steps pass compliance checks but the composed execution violates governing policy. Proposes a four-type taxonomy (Authority Creep, Threshold Laundering, Cumulative Sum Violation, Context Collapse) and a provenance-aware runtime architecture for trace-level policy evaluation.

Names a failure class that step-scoped monitors structurally cannot detect, with a taxonomy and runtime architecture for repair. Relevant to anyone governing agentic systems in regulated settings.

Cited by

No citations on record.