[
  {
    "id": "daplab-9-patterns",
    "name": "9 Critical Failure Patterns of Coding Agents",
    "owner": "Columbia DAPLab (Reya Vir and collaborators)",
    "version": "2026-01",
    "first_published": "2026-01-08",
    "url": "https://daplab.cs.columbia.edu/general/2026/01/08/9-critical-failure-patterns-of-coding-agents.html",
    "license": "Blog post; attribution requested",
    "scope": "Empirical failure patterns observed across 15+ applications and 5 state-of-the-art coding agents (Claude Code, Cline, Cursor, v0, Replit).\n",
    "categories": [
      {
        "external_id": "DAP-01",
        "label": "Incomplete Task Execution",
        "description": "Agent claims completion while key subtasks remain undone."
      },
      {
        "external_id": "DAP-02",
        "label": "Fabricated References",
        "description": "Agent emits code referencing non-existent functions or packages."
      },
      {
        "external_id": "DAP-03",
        "label": "Stale API Usage",
        "description": "Agent uses deprecated or removed APIs."
      },
      {
        "external_id": "DAP-04",
        "label": "Cascading Error Recovery",
        "description": "Agent's fix attempts introduce new errors that compound."
      },
      {
        "external_id": "DAP-05",
        "label": "Context Window Overflow",
        "description": "Important context is truncated; agent drifts."
      },
      {
        "external_id": "DAP-06",
        "label": "Tool Misinvocation",
        "description": "Correct tool, wrong arguments or wrong sequence."
      },
      {
        "external_id": "DAP-07",
        "label": "Silent Test-Passing Workaround",
        "description": "Agent modifies tests or test harness to mask failure."
      },
      {
        "external_id": "DAP-08",
        "label": "Destructive Action",
        "description": "Agent executes rm/force-push/commit bypass without human approval."
      },
      {
        "external_id": "DAP-09",
        "label": "Goal Drift",
        "description": "Agent's working goal diverges from the user's stated goal."
      }
    ],
    "notes": "Category labels paraphrased for seed. Verify against the original DAPLab post before v0.1 release. Reya Vir is the top co-author candidate identified by the landscape agent.\n",
    "inclusion_rationale": "Closest spiritual sibling to SAICA-KG — an empirical classification of coding-agent failures. Not a registry and not a KG, which is exactly the gap SAICA-KG fills. Crosswalks from DAPLab patterns to SAICA-KG FailureMode tags are a high-value day-1 artifact.\n"
  },
  {
    "id": "mast",
    "name": "MAST: Multi-Agent System Failure Taxonomy",
    "owner": "Cemri et al. (UC Berkeley + collaborators)",
    "version": "2025",
    "first_published": "2025-03-17",
    "url": "https://arxiv.org/abs/2503.13657",
    "license": "arXiv preprint; taxonomy content CC-BY",
    "paper_id": "cemri-2025-mast",
    "scope": "Failure modes of multi-agent LLM systems, grounded-theory derived from 1,600 annotated traces across 7 multi-agent frameworks.\n",
    "categories": [
      {
        "external_id": "MAST-FM-01",
        "label": "Role Violation",
        "description": "Agent acts outside its declared role or scope."
      },
      {
        "external_id": "MAST-FM-02",
        "label": "Task Decomposition Failure",
        "description": "Agent fails to break a task into solvable subtasks."
      },
      {
        "external_id": "MAST-FM-03",
        "label": "Coordination Failure",
        "description": "Agents fail to coordinate on shared state or handoff."
      },
      {
        "external_id": "MAST-FM-04",
        "label": "Redundant Work",
        "description": "Multiple agents duplicate effort without coordination."
      },
      {
        "external_id": "MAST-FM-05",
        "label": "Communication Breakdown",
        "description": "Messages between agents are lost, misinterpreted, or ignored."
      },
      {
        "external_id": "MAST-FM-06",
        "label": "Stalling / Deadlock",
        "description": "Agents block waiting on each other with no progress."
      },
      {
        "external_id": "MAST-FM-07",
        "label": "Premature Termination",
        "description": "Agent ends a task before completion criteria are met."
      },
      {
        "external_id": "MAST-FM-08",
        "label": "Goal Drift",
        "description": "Agent's working goal diverges from the user's stated goal."
      },
      {
        "external_id": "MAST-FM-09",
        "label": "Tool Misuse",
        "description": "Agent invokes an available tool incorrectly or unnecessarily."
      },
      {
        "external_id": "MAST-FM-10",
        "label": "Fabrication in Inter-Agent Messages",
        "description": "Agent sends factually incorrect information to another agent."
      },
      {
        "external_id": "MAST-FM-11",
        "label": "Verification Failure",
        "description": "Agent accepts incorrect output without checking."
      },
      {
        "external_id": "MAST-FM-12",
        "label": "Escalation Failure",
        "description": "Agent does not escalate to a human when appropriate."
      },
      {
        "external_id": "MAST-FM-13",
        "label": "Context Loss",
        "description": "Relevant prior context is dropped or truncated."
      },
      {
        "external_id": "MAST-FM-14",
        "label": "Infinite Loop",
        "description": "Agent cycles over the same state or tool call indefinitely."
      }
    ],
    "notes": "Category labels here are paraphrased for seed purposes and MUST be verified against the MAST paper's Table 1 before v0.1 release. Full-text review is a Week-1 task.\n",
    "inclusion_rationale": "MAST is the most-cited 2025 multi-agent failure taxonomy and is a primary ingestion target for SAICA-KG's crosswalk contribution. Its MAS focus complements SAICA-KG's single-agent coding-agent scope — crosswalks to FailureMode tags surface which modes generalize across agent topologies.\n"
  },
  {
    "id": "microsoft-airt-2025",
    "name": "Microsoft AI Red Team: Taxonomy of Failure Modes in Agentic AI Systems",
    "owner": "Microsoft AI Red Team (AIRT)",
    "version": "2025-04",
    "first_published": "2025-04-24",
    "url": "https://www.microsoft.com/en-us/security/blog/2025/04/24/new-whitepaper-outlines-the-taxonomy-of-failure-modes-in-ai-agents/",
    "license": "Microsoft whitepaper; attribution required",
    "scope": "Failure modes of agentic AI systems, split along two axes: safety vs. security (impact) and novel vs. existing (whether the mode is unique to agentic AI or inherited from prior generative AI). Single- vs. multi-agent distinctions are surfaced in descriptions. Derived from Microsoft AI Red Team's operational experience and a 30-page April 2025 whitepaper.\n",
    "categories": [
      {
        "external_id": "MS-NS-01",
        "label": "Agent Compromise",
        "description": "Unauthorized control or manipulation of an agent; one corrupted node can subvert entire multi-agent workflows.\n"
      },
      {
        "external_id": "MS-NS-02",
        "label": "Agent Injection",
        "description": "Insertion of malicious instructions into agent prompts, plans, or tool-call arguments."
      },
      {
        "external_id": "MS-NS-03",
        "label": "Agent Impersonation",
        "description": "Threat actors register counterfeit agents that appear legitimate, tricking other modules into leaking data or executing actions.\n"
      },
      {
        "external_id": "MS-NS-04",
        "label": "Agent Flow Manipulation",
        "description": "Unauthorized alteration of agent workflow execution paths, e.g., injecting tokens like STOP to reroute workflows or bypass safety logic.\n"
      },
      {
        "external_id": "MS-NS-05",
        "label": "Multi-Agent Jailbreaks",
        "description": "Jailbreak payload split across agent messages; recombination bypasses single-prompt detectors. Specific to multi-agent topologies.\n"
      },
      {
        "external_id": "MS-NF-01",
        "label": "Intra-Agent Responsible AI Concerns",
        "description": "Toxic or RAI-violating content propagates through internal inter-agent messages and reaches users or logs without filtering.\n"
      },
      {
        "external_id": "MS-NF-02",
        "label": "Bias in Resource Allocation",
        "description": "Discriminatory distribution of resources across users when agents act as allocators."
      },
      {
        "external_id": "MS-NF-03",
        "label": "Organizational Knowledge Degradation",
        "description": "Heavy automation erodes human understanding of the underlying system; recovery is difficult when agents fail.\n"
      },
      {
        "external_id": "MS-NF-04",
        "label": "Prioritization Risks (Performance Over Safety)",
        "description": "Goal-driven agents skip safety checks to optimize KPIs, prioritizing performance metrics over user safety.\n"
      },
      {
        "external_id": "MS-ES-01",
        "label": "Memory Poisoning",
        "description": "Attackers embed malicious instructions into long-term memory; every future recall re-executes the attack. Flagship case study in the paper.\n"
      },
      {
        "external_id": "MS-ES-02",
        "label": "Cross-Domain Prompt Injection (XPIA)",
        "description": "External content (PDFs, calendar invites, web pages) contains hidden commands that hijack agent decision logic.\n"
      },
      {
        "external_id": "MS-ES-03",
        "label": "Human-in-the-Loop Bypass",
        "description": "Circumvention of human oversight and approval gates."
      },
      {
        "external_id": "MS-ES-04",
        "label": "Incorrect Permissions Management",
        "description": "Over-scoped agent permissions let agents access files or perform actions beyond the user's intent.\n"
      },
      {
        "external_id": "MS-ES-05",
        "label": "Insufficient Isolation",
        "description": "Inadequate separation between agent operational boundaries, enabling lateral compromise."
      },
      {
        "external_id": "MS-EF-01",
        "label": "Bias Amplification",
        "description": "Magnification of existing prejudices in agent outputs across iterative loops."
      },
      {
        "external_id": "MS-EF-02",
        "label": "Hallucinations",
        "description": "Generation of false or fabricated information, including fabricated tool calls and references."
      },
      {
        "external_id": "MS-EF-03",
        "label": "Misinterpretation of Instructions",
        "description": "Incorrect understanding of user directives leading to unintended actions."
      },
      {
        "external_id": "MS-EF-04",
        "label": "Lack of Transparency",
        "description": "Insufficient clarity over agent behavior, preventing informed user consent and oversight."
      }
    ],
    "notes": "Category labels paraphrased from the Microsoft Security Blog summary and Adversa/MarkTechPost coverage; pending full-text verification against the 30-page whitepaper PDF. Full expansion (e.g., sub-modes under Memory Poisoning) and any additional categories is a v0.2 task.\n",
    "inclusion_rationale": "The Microsoft AIRT taxonomy is the most operationally-grounded agentic AI failure taxonomy from a frontier vendor and is widely cited in industry agentic-security discussions. Ingesting it as a first-class Taxonomy node gives SAICA-KG crosswalks into: supply_chain_attack (MS-NS-02, MS-ES-02), context_pollution (MS-ES-01 memory poisoning), scope_creep (MS-ES-04, MS-NS-04), fabrication (MS-EF-02), and security_vulnerability (most of the NS/ES axis). The novel-vs-existing axis also informs the single-agent-vs-multi-agent framing in SAICA-KG's Tool nodes.\n"
  },
  {
    "id": "owasp-agentic-top-10-2026",
    "name": "OWASP Top 10 for Agentic Applications 2026",
    "owner": "OWASP Gen AI Security Project",
    "version": "2026",
    "first_published": "2025-12-09",
    "url": "https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/",
    "license": "Open (OWASP), CC-BY",
    "scope": "The ten most critical security risks facing autonomous and agentic AI applications. Globally peer-reviewed, developed with 100+ industry experts and researchers. The most visible formal taxonomy of agentic AI risk.\n",
    "categories": [
      {
        "external_id": "ASI01",
        "label": "Agent Goal Hijack",
        "description": "Attacker-induced diversion of an agent from its user-assigned goal."
      },
      {
        "external_id": "ASI02",
        "label": "Tool Misuse and Exploitation",
        "description": "Abuse of the agent's tool surface to exceed intended authority."
      },
      {
        "external_id": "ASI03",
        "label": "Identity and Privilege Abuse",
        "description": "Misuse of agent identities or escalation of privileges."
      },
      {
        "external_id": "ASI04",
        "label": "Agentic Supply Chain Vulnerabilities",
        "description": "Compromise via the agent's software, tool, model, or data supply chain, including hallucinated-package-name typosquatting (slopsquatting).\n"
      },
      {
        "external_id": "ASI05",
        "label": "Unexpected Code Execution (RCE)",
        "description": "Agent emits or runs code producing unintended remote execution."
      },
      {
        "external_id": "ASI06",
        "label": "Memory & Context Poisoning",
        "description": "Attacker manipulation of agent memory or long-context state."
      },
      {
        "external_id": "ASI07",
        "label": "Insecure Inter-Agent Communication",
        "description": "Weakly authenticated or validated messages between agents."
      },
      {
        "external_id": "ASI08",
        "label": "Cascading Failures",
        "description": "Errors that propagate across agent subsystems or the agent fleet."
      },
      {
        "external_id": "ASI09",
        "label": "Human-Agent Trust Exploitation",
        "description": "Social-engineering of humans by or via an agent."
      },
      {
        "external_id": "ASI10",
        "label": "Rogue Agents",
        "description": "Unsanctioned or self-modifying agents operating outside policy."
      }
    ],
    "notes": "Full ASI descriptions will be expanded during v0.1 seeding; the category labels here are canonical from the OWASP Gen AI Security Project.\n",
    "inclusion_rationale": "OWASP Agentic Top 10 is the single most visible formal taxonomy of agentic AI risk. Ingesting it as a first-class Taxonomy node lets SAICA-KG crosswalk OWASP categories to its FailureMode facets and answer OWASP-keyed queries through the website and JSON API. This is the \"OWASP/ATT&CK for AI coding agent supervision\" positioning operationalized.\n"
  },
  {
    "id": "shah-2026-agentic-faults",
    "name": "Characterizing Faults in Agentic AI: A Taxonomy of Types, Symptoms, and Root Causes",
    "owner": "Shah et al. (arXiv:2603.06847)",
    "version": "2026-03",
    "first_published": "2026-03-06",
    "url": "https://arxiv.org/abs/2603.06847",
    "license": "arXiv preprint; CC-BY",
    "paper_id": "shah-2026-characterizing-faults",
    "scope": "Grounded-theory taxonomy of faults in agentic AI systems, derived from 13,602 issues and PRs across 40 open-source agentic-AI repositories, with stratified sampling to 385 faults and developer-study validation (n=145). Three parallel taxonomies: fault types (5 dimensions, 13 major categories, 37 fault subcategories); 13 symptom classes; 12 root-cause categories. This is the empirical corpus SAICA-KG uses for its v0.1 back-test.\n",
    "categories": [
      {
        "external_id": "SHAH-DIM-01",
        "label": "Agent Cognition and Orchestration",
        "description": "Failures in cognitive architecture and control flows unique to agentic systems (83 faults in the sample). Covers LLM integration and agent lifecycle/state.\n"
      },
      {
        "external_id": "SHAH-DIM-02",
        "label": "Tooling, Integration, and Actuation",
        "description": "Defects in tool invocation and external-system interaction (66 faults). Covers API usage, external connectivity, resource manipulation, coordination, and observability of tool calls.\n"
      },
      {
        "external_id": "SHAH-DIM-03",
        "label": "Perception, Context, and Memory",
        "description": "Faults in information ingestion, context preservation, and input interpretation (72 faults). Covers persistence and input-interpretation logic.\n"
      },
      {
        "external_id": "SHAH-DIM-04",
        "label": "Runtime and Environment Grounding",
        "description": "Execution environment, dependency stack, and platform-constraint mismatches (87 faults). The single largest fault dimension by count, dominated by dependency management.\n"
      },
      {
        "external_id": "SHAH-DIM-05",
        "label": "System Reliability and Observability",
        "description": "Robustness deficiencies and diagnostic challenges in autonomous systems (67 faults). Covers error recovery, UI/visualization, and documentation.\n"
      },
      {
        "external_id": "SHAH-MAJ-01",
        "label": "LLM Integration Faults",
        "description": "Faults in how the agent integrates with its LLM (45 faults).",
        "parent_id": "SHAH-DIM-01"
      },
      {
        "external_id": "SHAH-MAJ-02",
        "label": "Agent Lifecycle and State",
        "description": "Faults in agent startup, execution, termination, and state tracking (38 faults).",
        "parent_id": "SHAH-DIM-01"
      },
      {
        "external_id": "SHAH-MAJ-03",
        "label": "Tool Execution and API Usage",
        "description": "Incorrect or malformed tool/API invocations (17 faults).",
        "parent_id": "SHAH-DIM-02"
      },
      {
        "external_id": "SHAH-MAJ-04",
        "label": "External Connectivity and Access",
        "description": "Connection setup and authentication failures (19 faults).",
        "parent_id": "SHAH-DIM-02"
      },
      {
        "external_id": "SHAH-MAJ-05",
        "label": "Resource Manipulation",
        "description": "Resource-handling and database-misconfiguration faults (16 faults).",
        "parent_id": "SHAH-DIM-02"
      },
      {
        "external_id": "SHAH-MAJ-06",
        "label": "System Coordination",
        "description": "Synchronization errors across components or agents (6 faults).",
        "parent_id": "SHAH-DIM-02"
      },
      {
        "external_id": "SHAH-MAJ-07",
        "label": "System Observability",
        "description": "Logging or telemetry defects (8 faults).",
        "parent_id": "SHAH-DIM-02"
      },
      {
        "external_id": "SHAH-MAJ-08",
        "label": "Context and State Persistence",
        "description": "Memory-persistence and state-load/save failures (12 faults).",
        "parent_id": "SHAH-DIM-03"
      },
      {
        "external_id": "SHAH-MAJ-09",
        "label": "Input Interpretation and Logic",
        "description": "Type handling, validation, encoding, and logic/constraint violations (60 faults).",
        "parent_id": "SHAH-DIM-03"
      },
      {
        "external_id": "SHAH-MAJ-10",
        "label": "Dependency Management",
        "description": "Dependency specification, import resolution, installation, and package resolver/manager failures (67 faults). The single largest major category.\n",
        "parent_id": "SHAH-DIM-04"
      },
      {
        "external_id": "SHAH-MAJ-11",
        "label": "Platform Compatibility",
        "description": "Environment, platform, API, and integration compatibility mismatches (20 faults).",
        "parent_id": "SHAH-DIM-04"
      },
      {
        "external_id": "SHAH-MAJ-12",
        "label": "Robustness and Error Recovery",
        "description": "Exception-handling and implementation defects (49 faults).",
        "parent_id": "SHAH-DIM-05"
      },
      {
        "external_id": "SHAH-MAJ-13",
        "label": "UI, Visualization, and Documentation Defects",
        "description": "UI defects (13 faults) and documentation defects (5 faults).",
        "parent_id": "SHAH-DIM-05"
      },
      {
        "external_id": "SHAH-RC-01",
        "label": "Dependency and Integration Failures",
        "description": "Contract violations between probabilistic artifacts and deterministic interfaces (19.5%)."
      },
      {
        "external_id": "SHAH-RC-02",
        "label": "Data and Type Handling Failures",
        "description": "Type mismatches and validation gaps (17.6%)."
      },
      {
        "external_id": "SHAH-RC-03",
        "label": "Incorrect Token Management",
        "description": "Tracking and counting errors affecting context windows."
      },
      {
        "external_id": "SHAH-RC-04",
        "label": "Authentication and Credential Issues",
        "description": "Fragile token refresh and credential-expiration handling."
      },
      {
        "external_id": "SHAH-RC-05",
        "label": "Datetime Conversion Errors",
        "description": "Inappropriate temporal-value transformations."
      },
      {
        "external_id": "SHAH-RC-06",
        "label": "State Management Defects",
        "description": "Inconsistent internal state across interactions."
      },
      {
        "external_id": "SHAH-RC-07",
        "label": "Configuration Inconsistencies",
        "description": "Wrong defaults and misconfiguration."
      },
      {
        "external_id": "SHAH-RC-08",
        "label": "API Evolution and Schema Drift",
        "description": "Incompatibilities from upstream API changes."
      },
      {
        "external_id": "SHAH-RC-09",
        "label": "Environment-Level Issues",
        "description": "Platform-specific constraints and virtual-environment problems."
      },
      {
        "external_id": "SHAH-RC-10",
        "label": "Documentation Gaps",
        "description": "Missing or incorrect configuration guidance."
      },
      {
        "external_id": "SHAH-RC-11",
        "label": "Exception Suppression",
        "description": "Swallowed errors reducing observability."
      },
      {
        "external_id": "SHAH-RC-12",
        "label": "Ecosystem Fragility",
        "description": "Rapid upstream evolution creating tight-coupling vulnerabilities."
      }
    ],
    "notes": "Top-level structure (5 dimensions, 13 major categories, 12 root causes) is canonical per the arXiv v1 HTML; the 37 fault subcategories are enumerated in the paper and will be expanded into full category rows as a v0.2 task. Symptom-class and full subcategory labels are pending full-text verification. This taxonomy is the empirical back-test corpus: SAICA-KG's FailureMode axis will be evaluated against whether SHAH-* faults map cleanly into its eight-class FailureMode enum.\n",
    "inclusion_rationale": "Shah et al. 2026 is the single most comprehensive empirical fault characterization for agentic AI to date, with a 385-fault grounded-theory sample across 40 real repos — exactly the scale SAICA-KG needs to argue its FailureMode partition is honest-MECE. Intended crosswalks: SHAH-MAJ-10 (Dependency Management) -> obsolescence + supply_chain_attack; SHAH-MAJ-03 (Tool Execution/API Usage) -> logic_error + fabrication (API); SHAH-MAJ-08 (Context/State Persistence) -> context_pollution; SHAH-MAJ-09 (Input Interpretation/Logic) -> logic_error; SHAH-RC-04 (Auth/Credential) -> security_vulnerability; SHAH-RC-08 (API Schema Drift) -> obsolescence. Coverage ratio against SAICA-KG's FailureMode enum is the v0.1 honest-MECE evidence.\n"
  },
  {
    "id": "swiss-cheese-model",
    "name": "Swiss Cheese Model for AI Safety: Multi-Layered Runtime Guardrail Taxonomy",
    "owner": "Shamsujjoha, Lu, Zhao, Zhu (CSIRO Data61)",
    "version": "2025-01 (v3)",
    "first_published": "2024-08-05",
    "last_updated": "2025-01-27",
    "url": "https://arxiv.org/abs/2408.02205",
    "license": "arXiv preprint; CC-BY",
    "paper_id": "shamsujjoha-2024-swiss-cheese-model",
    "scope": "A SLR-derived taxonomy of multi-layered runtime guardrails for foundation-model-based agents. Organized along three architectural dimensions (pipelines, artifacts, actions) plus quality attributes that guardrails should satisfy. Evoked as a \"Swiss Cheese\" defense-in-depth reference architecture for AI safety by design.\n",
    "categories": [
      {
        "external_id": "SC-QA-01",
        "label": "Accuracy",
        "description": "Mitigates hallucinations, misinformation, and disinformation to ensure factually correct outputs.",
        "parent_id": "SC-QA"
      },
      {
        "external_id": "SC-QA-02",
        "label": "Efficiency",
        "description": "Prevents resource-intensive tasks and endless loops; manages cost through dynamic resource management.",
        "parent_id": "SC-QA"
      },
      {
        "external_id": "SC-QA-03",
        "label": "Privacy",
        "description": "Protects against sensitive-data leakage via direct responses or statistical inference.",
        "parent_id": "SC-QA"
      },
      {
        "external_id": "SC-QA-04",
        "label": "Security",
        "description": "Defends against malicious activities and adversarial attacks; preserves data confidentiality and integrity.",
        "parent_id": "SC-QA"
      },
      {
        "external_id": "SC-QA-05",
        "label": "Safety",
        "description": "Prevents harmful or misleading outputs, especially in high-stakes domains.",
        "parent_id": "SC-QA"
      },
      {
        "external_id": "SC-QA-06",
        "label": "Fairness",
        "description": "Addresses bias and discrimination from training data, algorithms, or deployment.",
        "parent_id": "SC-QA"
      },
      {
        "external_id": "SC-QA-07",
        "label": "Compliance",
        "description": "Ensures adherence to legal, regulatory, and industry standards via continuous monitoring.",
        "parent_id": "SC-QA"
      },
      {
        "external_id": "SC-QA-08",
        "label": "Generalizability",
        "description": "Functions effectively across multiple layers and diverse scenarios without per-case configuration.",
        "parent_id": "SC-QA"
      },
      {
        "external_id": "SC-QA-09",
        "label": "Customizability",
        "description": "Tailored protection for specific requirements at different operational layers.",
        "parent_id": "SC-QA"
      },
      {
        "external_id": "SC-QA-10",
        "label": "Adaptability",
        "description": "Adjusts dynamically to varying conditions and emerging threats without manual reconfiguration.",
        "parent_id": "SC-QA"
      },
      {
        "external_id": "SC-QA-11",
        "label": "Traceability",
        "description": "Detailed logs of decision paths, inputs, outputs, and data sources for auditing.",
        "parent_id": "SC-QA"
      },
      {
        "external_id": "SC-QA-12",
        "label": "Portability",
        "description": "Guardrail transfers across FM-based agents with minimal adjustment.",
        "parent_id": "SC-QA"
      },
      {
        "external_id": "SC-QA-13",
        "label": "Interoperability",
        "description": "Functions seamlessly across heterogeneous agents, technologies, and interfaces.",
        "parent_id": "SC-QA"
      },
      {
        "external_id": "SC-QA-14",
        "label": "Interpretability",
        "description": "Provides clarity on how a guardrail operates, enabling user understanding and trust.",
        "parent_id": "SC-QA"
      },
      {
        "external_id": "SC-PIPE-01",
        "label": "Prompt Stage",
        "description": "Incoming user queries analyzed for harmful content and safety alignment before processing.",
        "parent_id": "SC-PIPE"
      },
      {
        "external_id": "SC-PIPE-02",
        "label": "Intermediate Results Stage",
        "description": "Outputs at workflow stages verified for accuracy and safety before final generation.",
        "parent_id": "SC-PIPE"
      },
      {
        "external_id": "SC-PIPE-03",
        "label": "Final Results Stage",
        "description": "Final agent outputs checked for goal alignment and governance compliance.",
        "parent_id": "SC-PIPE"
      },
      {
        "external_id": "SC-ART-01",
        "label": "Goals",
        "description": "Declared user or task objectives.",
        "parent_id": "SC-ART"
      },
      {
        "external_id": "SC-ART-02",
        "label": "Context",
        "description": "Conversation and session context.",
        "parent_id": "SC-ART"
      },
      {
        "external_id": "SC-ART-03",
        "label": "Memory",
        "description": "Long-term or episodic agent memory.",
        "parent_id": "SC-ART"
      },
      {
        "external_id": "SC-ART-04",
        "label": "Reasoning",
        "description": "Chain-of-thought and intermediate reasoning traces.",
        "parent_id": "SC-ART"
      },
      {
        "external_id": "SC-ART-05",
        "label": "Plans",
        "description": "Multi-step action plans produced by the agent.",
        "parent_id": "SC-ART"
      },
      {
        "external_id": "SC-ART-06",
        "label": "Workflows",
        "description": "End-to-end orchestrated task flows.",
        "parent_id": "SC-ART"
      },
      {
        "external_id": "SC-ART-07",
        "label": "Tools",
        "description": "External callable tools and APIs.",
        "parent_id": "SC-ART"
      },
      {
        "external_id": "SC-ART-08",
        "label": "Knowledge Bases",
        "description": "Retrieval corpora and structured knowledge stores.",
        "parent_id": "SC-ART"
      },
      {
        "external_id": "SC-ART-09",
        "label": "Other Agents",
        "description": "Peer or subordinate agents in a multi-agent system.",
        "parent_id": "SC-ART"
      },
      {
        "external_id": "SC-ART-10",
        "label": "Foundation Models",
        "description": "The underlying LLM / FM itself.",
        "parent_id": "SC-ART"
      },
      {
        "external_id": "SC-ACT-01",
        "label": "Block",
        "description": "Prevent the action from executing.",
        "parent_id": "SC-ACT"
      },
      {
        "external_id": "SC-ACT-02",
        "label": "Filter",
        "description": "Strip disallowed content before forwarding.",
        "parent_id": "SC-ACT"
      },
      {
        "external_id": "SC-ACT-03",
        "label": "Flag",
        "description": "Emit a warning for downstream review.",
        "parent_id": "SC-ACT"
      },
      {
        "external_id": "SC-ACT-04",
        "label": "Modify",
        "description": "Transform the action or output to a safer variant.",
        "parent_id": "SC-ACT"
      },
      {
        "external_id": "SC-ACT-05",
        "label": "Validate",
        "description": "Check against a spec or schema before proceeding.",
        "parent_id": "SC-ACT"
      },
      {
        "external_id": "SC-ACT-06",
        "label": "Parallel Calls",
        "description": "Run multiple checkers concurrently for consensus.",
        "parent_id": "SC-ACT"
      },
      {
        "external_id": "SC-ACT-07",
        "label": "Retry",
        "description": "Regenerate with adjusted parameters.",
        "parent_id": "SC-ACT"
      },
      {
        "external_id": "SC-ACT-08",
        "label": "Fall Back",
        "description": "Switch to a safer alternative pathway.",
        "parent_id": "SC-ACT"
      },
      {
        "external_id": "SC-ACT-09",
        "label": "Human Intervention",
        "description": "Escalate to a human-in-the-loop.",
        "parent_id": "SC-ACT"
      },
      {
        "external_id": "SC-ACT-10",
        "label": "Defer",
        "description": "Delay the action until conditions are met.",
        "parent_id": "SC-ACT"
      },
      {
        "external_id": "SC-ACT-11",
        "label": "Isolate",
        "description": "Run the action in a sandbox or quarantined environment.",
        "parent_id": "SC-ACT"
      },
      {
        "external_id": "SC-ACT-12",
        "label": "Redundancy",
        "description": "Duplicate critical checks for defense in depth.",
        "parent_id": "SC-ACT"
      },
      {
        "external_id": "SC-ACT-13",
        "label": "Evaluate",
        "description": "Score against metrics before acting.",
        "parent_id": "SC-ACT"
      }
    ],
    "notes": "Category labels derived from the arXiv v3 HTML; quality-attribute and action lists are canonical. Parent_id groupings (SC-QA, SC-PIPE, SC-ART, SC-ACT) are synthetic grouping anchors used for hierarchy — the underlying paper treats them as dimensions rather than named parents. Pending full-text verification against the final ICSA 2025 version, including the Rules, Sources, Modality, and Underlying-Models dimensions which are deferred to v0.2 expansion.\n",
    "inclusion_rationale": "The Swiss Cheese Model paper gives SAICA-KG's Tool ontology an external architectural anchor for defense-in-depth reasoning: Tool nodes can be mapped to (pipeline stage x artifact x action) triples and evaluated against the 14 quality attributes. Strong crosswalks: SC-PIPE-01/02/03 aligns directly with SAICA-KG's TemporalPhase partition; SC-ART-* overlaps with LocusOfControl; SC-ACT-* overlaps with ControlParadigm (Block -> prevention, Flag -> detection, Modify/Retry -> correction, Fall-back/Isolate -> recovery). This makes the Swiss Cheese paper a first-class validation anchor for SAICA-KG's axis choices.\n"
  }
]