{
  "schema": "shareplane-evidence-register-v1",
  "artifactId": "artifact:let-intelligence-be-messy-make-authority-boring",
  "sources": [
    {
      "id": "src:nist-airmf-core",
      "organization": "NIST",
      "title": "AI Risk Management Framework Core",
      "url": "https://airc.nist.gov/airmf-resources/airmf/5-sec-core/",
      "retrieved": "2026-09-23",
      "reviewed": "2026-09-23",
      "quality": "authoritative-institutional-source",
      "supports": [
        "claim:lifecycle-risk",
        "claim:independent-assessment",
        "claim:production-monitoring",
        "claim:third-party-components"
      ],
      "doesNotEstablish": "Does not establish the article's original synthesis that authority should be the deterministic boundary around probabilistic semantics."
    },
    {
      "id": "src:nist-airmf-gaip",
      "organization": "NIST",
      "title": "Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile",
      "url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence",
      "retrieved": "2026-09-23",
      "reviewed": "2026-09-23",
      "quality": "authoritative-institutional-source",
      "supports": [
        "claim:system-lifecycle-risk"
      ],
      "doesNotEstablish": "Does not validate any specific GhostMesh implementation or claim that deterministic controls eliminate AI risk."
    },
    {
      "id": "src:owasp-excessive-agency",
      "organization": "OWASP GenAI Security Project",
      "title": "LLM06:2025 Excessive Agency",
      "url": "https://genai.owasp.org/llmrisk/llm062025-excessive-agency/",
      "retrieved": "2026-09-23",
      "reviewed": "2026-09-23",
      "quality": "authoritative-institutional-source",
      "supports": [
        "claim:excessive-agency",
        "claim:least-privilege",
        "claim:human-approval"
      ],
      "doesNotEstablish": "Does not prove that all agent failures are caused by excessive permissions or autonomy."
    },
    {
      "id": "src:cisa-secure-by-design",
      "organization": "CISA and international partners",
      "title": "Shifting the Balance of Cybersecurity Risk: Principles and Approaches for Security-by-Design and -Default",
      "url": "https://www.cisa.gov/sites/default/files/2023-06/principles_approaches_for_security-by-design-default_508c.pdf",
      "retrieved": "2026-09-23",
      "reviewed": "2026-09-23",
      "quality": "authoritative-institutional-source",
      "supports": [
        "claim:defense-in-depth",
        "claim:containment",
        "claim:least-privilege"
      ],
      "doesNotEstablish": "Does not address AI-specific reward hacking or prove unknown failures can be made harmless."
    },
    {
      "id": "src:anthropic-reward-seeker",
      "organization": "Anthropic",
      "title": "Training a Misaligned Reward Seeker",
      "url": "https://alignment.anthropic.com/2026/reward-seeker/",
      "retrieved": "2026-09-23",
      "reviewed": "2026-09-23",
      "quality": "primary-research",
      "supports": [
        "claim:reward-hacking",
        "claim:unanticipated-reward-hacks"
      ],
      "doesNotEstablish": "Research result from a specific experimental training setup. It does not establish that deployed agents generally act deceptively or intentionally."
    },
    {
      "id": "src:anthropic-practices-2026",
      "organization": "Anthropic",
      "title": "Improving our alignment and security practices",
      "url": "https://www.anthropic.com/news/improving-alignment-security-efforts",
      "retrieved": "2026-09-23",
      "reviewed": "2026-09-23",
      "quality": "first-party-engineering-report",
      "supports": [
        "claim:production-reward-hacking-observation",
        "claim:monitoring-limits"
      ],
      "doesNotEstablish": "First-party report about Anthropic systems. It does not generalize by itself to every model provider or deployment."
    },
    {
      "id": "src:openai-agent-safety",
      "organization": "OpenAI",
      "title": "Safety in building agents",
      "url": "https://developers.openai.com/api/docs/guides/agent-builder-safety",
      "retrieved": "2026-09-23",
      "reviewed": "2026-09-23",
      "quality": "official-documentation",
      "supports": [
        "claim:layered-agent-controls",
        "claim:tool-approval",
        "claim:structured-boundaries"
      ],
      "doesNotEstablish": "Product guidance is not independent proof that any specific agent deployment is safe."
    }
  ],
  "claims": [
    {
      "id": "claim:system-lifecycle-risk",
      "posture": "EXTERNALLY_EVIDENCED",
      "text": "AI trustworthiness is a system and lifecycle concern rather than a property established by one model behavior test.",
      "sources": [
        "src:nist-airmf-core",
        "src:nist-airmf-gaip"
      ]
    },
    {
      "id": "claim:independent-assessment",
      "posture": "EXTERNALLY_EVIDENCED",
      "text": "Independent assessment and production monitoring are established AI risk-management practices.",
      "sources": [
        "src:nist-airmf-core"
      ]
    },
    {
      "id": "claim:excessive-agency",
      "posture": "EXTERNALLY_EVIDENCED",
      "text": "Excessive functionality, permissions, and autonomy materially increase agentic risk.",
      "sources": [
        "src:owasp-excessive-agency"
      ]
    },
    {
      "id": "claim:defense-in-depth",
      "posture": "EXTERNALLY_EVIDENCED",
      "text": "Defense in depth, least privilege, sandboxing, and segregation reduce the impact of individual control failures.",
      "sources": [
        "src:cisa-secure-by-design",
        "src:owasp-excessive-agency"
      ]
    },
    {
      "id": "claim:reward-hacking",
      "posture": "EXTERNALLY_EVIDENCED",
      "text": "Reward hacking occurs in reinforcement-learning settings and can involve behaviors not anticipated by environment designers.",
      "sources": [
        "src:anthropic-reward-seeker",
        "src:anthropic-practices-2026"
      ]
    },
    {
      "id": "claim:layered-agent-controls",
      "posture": "EXTERNALLY_EVIDENCED",
      "text": "Current agent safety guidance uses layered controls such as approvals, structured boundaries, guardrails, isolation, tracing, and evaluations rather than relying on one safeguard.",
      "sources": [
        "src:openai-agent-safety",
        "src:owasp-excessive-agency"
      ]
    },
    {
      "id": "claim:probabilistic-semantics-deterministic-authority",
      "posture": "ORIGINAL_SYNTHESIS",
      "text": "Probabilistic semantics should be allowed to explore while consequential authority is governed by deterministic, independently enforced boundaries.",
      "sources": []
    },
    {
      "id": "claim:trust-without-trusting-components",
      "posture": "ORIGINAL_SYNTHESIS",
      "text": "A trustworthy system can be designed so no single model, software component, operator, or validator must be trusted unconditionally.",
      "sources": []
    },
    {
      "id": "claim:unknown-unknown-containment",
      "posture": "RECOMMENDATION",
      "text": "Architecture should prioritize invariants and containment that bound unknown failures rather than attempting to enumerate every possible bad behavior.",
      "sources": [
        "src:cisa-secure-by-design",
        "src:nist-airmf-core"
      ]
    },
    {
      "id": "claim:owner-observation",
      "posture": "OWNER_OBSERVATION",
      "text": "Sustained use of coding agents has shown failures including stale-evidence pursuit, incorrect assumptions, and wrong-objective optimization that are operationally manageable when authority remains bounded.",
      "sources": []
    }
  ]
}
