Datasets:

validation
/

ai-validation-checklists

+# Data Dictionary
+## Controlled values
+### system_type
+- model
+- agent
+- rag
+- tool
+- data
+- multimodal
+- system
+### severity
+- medium
+- high
+- critical
+### lifecycle_stage
+- design
+- pre-deployment
+- deployment
+- monitoring
+### automation_level
+- manual
+- semi-automated
+- automated
+## Interpretation
+`threshold_guidance` provides a starting point for project-specific validation design. It is not a universal compliance threshold.
+`source_type = framework-inspired` means the checklist item is an independent operationalization informed by the cited source rather than a verbatim source requirement.

README.md ADDED Viewed

	@@ -0,0 +1,231 @@

+---
+pretty_name: AI Validation Checklists
+language:
+- en
+tags:
+- ai
+- validation
+- evaluation
+- agents
+- rag
+- multimodal
+- safety
+- reliability
+- governance
+task_categories:
+- text-classification
+---
+# AI Validation Checklists
+**AI Validation Checklists** is a structured reference dataset for validating AI systems across multiple system types and lifecycle stages.
+It is designed for practical use in:
+- model validation,
+- agent validation,
+- RAG validation,
+- tool-use validation,
+- data validation,
+- multimodal validation,
+- system-level validation and governance.
+The dataset currently contains **60 practical validation checks**.
+> **Important:** These checks are an independent, practical framework authored for this dataset. They are **not verbatim requirements**, certifications, legal advice, or an official checklist from NIST, ISO, OWASP, Hugging Face, or any other cited organization.
+## System types
+| System type | Checks |
+|---|---:|
+| Model | 10 |
+| Agent | 10 |
+| RAG | 10 |
+| Tool | 8 |
+| Data | 8 |
+| Multimodal | 8 |
+| System | 6 |
+## Schema
+| Field | Description |
+|---|---|
+| `id` | Stable validation-check identifier |
+| `system_type` | Model, agent, RAG, tool, data, multimodal, or system |
+| `validation_area` | Validation topic such as robustness, grounding, permissions, provenance, or monitoring |
+| `check` | Concrete validation question |
+| `expected_evidence` | Evidence that can support the validation activity |
+| `metric` | Example metric or observable |
+| `threshold_guidance` | Practical guidance for defining a project-specific threshold |
+| `severity` | `medium`, `high`, or `critical` |
+| `lifecycle_stage` | Design, pre-deployment, deployment, or monitoring |
+| `automation_level` | Manual, semi-automated, or automated |
+| `source_type` | How the cited source relates to the check |
+| `source_name` | Reference source |
+| `source_url` | Primary/reference URL |
+| `notes` | Scope and interpretation notes |
+## Example
+```json
+{
+  "id": "VAL-AGENT-001",
+  "system_type": "agent",
+  "validation_area": "goal_completion",
+  "check": "Does the agent complete representative end-to-end tasks successfully?",
+  "expected_evidence": "Scenario suite, traces, final outcomes",
+  "metric": "Task success rate",
+  "threshold_guidance": "Set a minimum task success rate per workflow and risk level.",
+  "severity": "high",
+  "lifecycle_stage": "pre-deployment",
+  "automation_level": "automated",
+  "source_type": "framework-inspired",
+  "source_name": "NIST AI Risk Management Framework (AI RMF 1.0)",
+  "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10",
+  "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."
+}
+```
+## How to use the dataset
+A validation program can filter checks by:
+```text
+system_type
++
+validation_area
++
+severity
++
+lifecycle_stage
++
+automation_level
+```
+For example:
+- an agent team can select all `agent` checks with `high` or `critical` severity;
+- a RAG team can focus on `grounding`, `citation_accuracy`, `access_control`, and `poisoning`;
+- an enterprise validation review can combine `model`, `data`, and `system` checks before deployment;
+- a monitoring program can select checks whose lifecycle stage is `monitoring`.
+## Thresholds are not universal
+The `threshold_guidance` field intentionally avoids pretending that one numerical threshold works for every system.
+Validation thresholds should depend on:
+- intended use,
+- risk and impact,
+- user population,
+- operating environment,
+- baseline performance,
+- regulatory or contractual requirements,
+- business tolerance,
+- human oversight.
+Some controls — such as unauthorized privileged actions — may reasonably require **zero tolerated successes in the validation suite**, while performance metrics usually require use-case-specific targets.
+## Source methodology
+The dataset is informed by public resources covering AI risk management, evaluation, security, validation, governance, and testing.
+Primary references include:
+- **NIST AI Risk Management Framework (AI RMF 1.0)**
+  https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10
+- **NIST AI RMF: Generative AI Profile (NIST AI 600-1)**
+  https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence
+- **NIST SP 800-218A — Secure Software Development Practices for Generative AI**
+  https://www.nist.gov/publications/secure-software-development-practices-generative-ai-and-dual-use-foundation-models-ssdf
+- **OWASP Top 10 for LLM Applications 2025**
+  https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/
+- **ISO/IEC 42001:2023 — AI management systems**
+  https://www.iso.org/standard/42001
+- **Hugging Face Evaluate**
+  https://huggingface.co/docs/evaluate/index
+The checklist items are **editorial operationalizations inspired by these sources**. A citation means the source is relevant context for the check; it does not mean the source uses the same wording or defines the included metric/threshold.
+## Files
+- `validation_checklists.csv` — convenient for the Hugging Face Dataset Viewer and spreadsheet-style inspection
+- `validation_checklists.jsonl` — convenient for programmatic use and downstream applications
+## Suggested subsets for future versions
+A future release could split or expose views such as:
+```text
+models
+agents
+rag
+tools
+data
+multimodal
+enterprise
+```
+Additional future fields could include:
+- `control_family`
+- `test_method`
+- `example_test_case`
+- `risk_if_failed`
+- `required_role`
+- `framework_mapping`
+- `last_reviewed`
+## Quality principles
+### Practical, not performative
+Every row should translate into a test, review, evidence request, or operational check.
+### Evidence-oriented
+The dataset asks what evidence should exist, not only whether a system “seems safe.”
+### System-aware
+Models, agents, RAG systems, tools, data pipelines, and multimodal systems have different failure modes.
+### Lifecycle-aware
+Validation is not only a pre-launch activity. Some properties must be monitored and revalidated after deployment.
+### Framework-informed, framework-independent
+The dataset draws on established public resources without claiming to reproduce or replace them.
+## Contributing
+Useful contributions include:
+- additional validation checks,
+- improved metrics,
+- clearer evidence requirements,
+- new system types,
+- better test methods,
+- additional primary sources,
+- corrections to outdated references.
+Please keep proposed checks specific, testable, and source-aware.
+## Collaboration
+Collaboration around AI validation, agent reliability, evaluation, testing, governance, safety and validation infrastructure is welcome.
+**Contact:** agenten@magenta.de
+## Disclaimer
+This dataset is an independent technical resource.
+It is not an official NIST, ISO, OWASP, Hugging Face, regulatory, certification, or audit artifact. It does not provide legal, compliance, certification, or professional audit advice.
+Organizations should define validation criteria based on their own systems, risks, jurisdictions, contractual obligations and qualified professional guidance.
+---
+**Validation — evidence before deployment, monitoring after release.**

validation_checklists.csv ADDED Viewed

	@@ -0,0 +1,61 @@

+id,system_type,validation_area,check,expected_evidence,metric,threshold_guidance,severity,lifecycle_stage,automation_level,source_type,source_name,source_url,notes
+VAL-MODEL-001,model,task_performance,Does the model meet the minimum quality required for its intended task?,"Task-specific evaluation set, baseline model, evaluation report",Primary task metric,Define a use-case-specific minimum and compare against an approved baseline.,high,pre-deployment,automated,framework-inspired,Hugging Face — Choosing a metric,https://huggingface.co/docs/evaluate/en/choosing_a_metric,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-MODEL-002,model,robustness,Does model performance remain acceptably stable under realistic prompt or input variation?,"Perturbation suite, repeated runs, robustness report",Performance delta under perturbation,Define the maximum acceptable degradation for representative variations.,high,pre-deployment,automated,framework-inspired,NIST AI Risk Management Framework (AI RMF 1.0),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-MODEL-003,model,calibration,Are confidence signals or uncertainty indicators meaningfully aligned with observed correctness where they are used?,"Calibration dataset, confidence outputs, reliability analysis",Calibration error / selective accuracy,Set a calibration target appropriate to the decision context; do not use confidence without validation.,medium,pre-deployment,automated,framework-inspired,NIST AI Risk Management Framework (AI RMF 1.0),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-MODEL-004,model,failure_modes,Are known high-impact failure modes documented and reproducible in testing?,"Failure catalog, test prompts, incident examples, model card",Failure reproduction coverage,All identified high-impact failure modes should have at least one reproducible test.,high,pre-deployment,semi-automated,framework-inspired,NIST AI RMF: Generative AI Profile (NIST AI 600-1),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-MODEL-005,model,bias_fairness,"Has performance been evaluated across relevant groups, languages, domains or other meaningful slices?","Slice definitions, evaluation results, disparity analysis",Worst-slice performance / disparity,"Define acceptable disparity based on context, impact and population; document unsupported slices.",high,pre-deployment,automated,framework-inspired,NIST AI Risk Management Framework (AI RMF 1.0),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-MODEL-006,model,safety,Does the model resist known harmful or disallowed behavior within the intended deployment context?,"Safety test suite, red-team results, refusal/behavior logs",Safety violation rate,Set scenario-specific limits; critical prohibited behaviors may require zero tolerated occurrences in the test set.,critical,pre-deployment,semi-automated,framework-inspired,NIST AI RMF: Generative AI Profile (NIST AI 600-1),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-MODEL-007,model,versioning,"Can every deployed model instance be traced to an exact model, revision and configuration?","Model registry record, commit/revision ID, deployment manifest",Traceability coverage,100% of production deployments should map to an exact approved model revision.,high,deployment,automated,framework-inspired,NIST SP 800-218A — Secure Software Development Practices for Generative AI,https://www.nist.gov/publications/secure-software-development-practices-generative-ai-and-dual-use-foundation-models-ssdf,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-MODEL-008,model,regression,Are model updates regression-tested against previously approved capabilities and risks?,"Regression suite, prior baseline, change log",Regression pass rate / metric deltas,No release should proceed with unexplained high-impact regressions.,high,pre-deployment,automated,framework-inspired,NIST AI Risk Management Framework (AI RMF 1.0),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-MODEL-009,model,resource_behavior,"Are latency, throughput and memory behavior measured under representative load?","Load-test results, hardware profile, serving logs",p95 latency / throughput / memory,Define operational SLOs for the intended environment and validate under expected concurrency.,medium,pre-deployment,automated,framework-inspired,NIST AI Risk Management Framework (AI RMF 1.0),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-MODEL-010,model,documentation,"Are intended use, limitations, evaluation scope and unsupported conditions documented?","Model card, validation report, risk documentation",Documentation completeness,All deployment-critical limitations and validation boundaries should be documented before release.,high,pre-deployment,manual,framework-inspired,NIST AI Risk Management Framework (AI RMF 1.0),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-AGENT-001,agent,goal_completion,Does the agent complete representative end-to-end tasks successfully?,"Scenario suite, traces, final outcomes",Task success rate,Set a minimum task success rate per workflow and risk level.,high,pre-deployment,automated,framework-inspired,NIST AI Risk Management Framework (AI RMF 1.0),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-AGENT-002,agent,tool_selection,Does the agent select only appropriate and permitted tools for each task?,"Tool-call traces, policy definitions, scenario tests",Tool selection accuracy / policy violation rate,Unauthorized or explicitly disallowed tool selection should be zero in the validation suite.,critical,pre-deployment,automated,framework-inspired,OWASP Top 10 for LLM Applications 2025,https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-AGENT-003,agent,permissions,Are agent actions constrained by least-privilege permissions?,"Permission matrix, service-account scopes, execution logs",Unauthorized action rate,Zero successful actions outside the agent's approved permission scope.,critical,pre-deployment,semi-automated,framework-inspired,OWASP Top 10 for LLM Applications 2025,https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-AGENT-004,agent,trajectory,"Can important agent decisions, tool calls and observations be reconstructed from traces?","Structured traces, tool logs, timestamps, model/version metadata",Trace completeness,100% of high-impact actions should have reconstructable decision and execution evidence.,high,deployment,automated,framework-inspired,NIST AI Risk Management Framework (AI RMF 1.0),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-AGENT-005,agent,recovery,"Does the agent recover safely from tool failures, unavailable dependencies and malformed results?","Fault-injection scenarios, traces, recovery outcomes",Safe recovery rate,Define acceptable recovery behavior; no unsafe fallback actions in critical scenarios.,high,pre-deployment,automated,framework-inspired,NIST AI Risk Management Framework (AI RMF 1.0),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-AGENT-006,agent,termination,Does the agent stop when its goal is reached or when configured limits are exceeded?,"Loop tests, max-step settings, execution logs",Unbounded-loop rate / excess-step rate,"Zero unbounded executions; enforce hard runtime, step or cost limits.",high,pre-deployment,automated,framework-inspired,OWASP Top 10 for LLM Applications 2025,https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-AGENT-007,agent,human_oversight,Are high-impact actions escalated for human approval where required?,"Approval policy, approval logs, action traces",Approval bypass rate,Zero execution of actions that policy marks as requiring approval without recorded approval.,critical,deployment,automated,framework-inspired,NIST AI Risk Management Framework (AI RMF 1.0),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-AGENT-008,agent,prompt_injection,Does the agent resist instructions in untrusted content that attempt to override system or tool-use policies?,"Injection test corpus, retrieval/tool scenarios, traces",Injection success rate,Set a strict threshold based on impact; privileged action compromise should be zero in the test suite.,critical,pre-deployment,semi-automated,framework-inspired,OWASP Top 10 for LLM Applications 2025,https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-AGENT-009,agent,cost_control,"Are token, tool, compute and external-service costs bounded and observable?","Budgets, cost telemetry, run logs",Cost per successful task / budget breach rate,Define workflow budgets and require zero uncontrolled budget overruns.,medium,deployment,automated,framework-inspired,OWASP Top 10 for LLM Applications 2025,https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-AGENT-010,agent,delegation,"In multi-agent systems, are delegation boundaries and responsibility transfers explicit and traceable?","Agent graph, delegation events, scopes, traces",Delegation trace coverage,All inter-agent delegation affecting high-impact actions should be attributable and reconstructable.,high,pre-deployment,semi-automated,framework-inspired,NIST AI Risk Management Framework (AI RMF 1.0),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-RAG-001,rag,retrieval_relevance,Does retrieval return information relevant to the user question or task?,"Query set, retrieved passages, relevance judgments",Recall@k / nDCG / precision@k,Choose a retrieval metric and target based on corpus size and downstream task.,high,pre-deployment,automated,framework-inspired,Hugging Face — Choosing a metric,https://huggingface.co/docs/evaluate/en/choosing_a_metric,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-RAG-002,rag,grounding,Are generated factual claims supported by retrieved evidence when grounding is required?,"Answers, retrieved context, claim-evidence labels",Grounding error rate / supported-claim rate,Define a maximum unsupported-claim rate; stricter thresholds for high-impact use cases.,critical,pre-deployment,semi-automated,framework-inspired,NIST AI RMF: Generative AI Profile (NIST AI 600-1),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-RAG-003,rag,citation_accuracy,Do citations point to sources that actually support the associated claims?,"Generated citations, source documents, citation labels",Citation precision,Set a minimum citation-support rate; critical claims should require directly supporting evidence.,high,pre-deployment,semi-automated,framework-inspired,NIST AI RMF: Generative AI Profile (NIST AI 600-1),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-RAG-004,rag,source_provenance,Can each indexed document be traced to an approved source and version?,"Corpus manifest, source URLs, document hashes, ingestion logs",Provenance coverage,100% of production documents should have source and version metadata.,high,deployment,automated,framework-inspired,NIST AI Risk Management Framework (AI RMF 1.0),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-RAG-005,rag,access_control,Does retrieval enforce document-level access restrictions for the requesting user or agent?,"ACLs, test identities, retrieval logs",Unauthorized retrieval rate,Zero retrieval of documents outside the requester's approved access scope.,critical,pre-deployment,automated,framework-inspired,OWASP Top 10 for LLM Applications 2025,https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-RAG-006,rag,staleness,Are time-sensitive documents refreshed or retired according to defined freshness rules?,"Document timestamps, refresh policy, index audit",Stale-document rate,Define freshness windows per source class and alert on expired content.,medium,monitoring,automated,framework-inspired,NIST AI Risk Management Framework (AI RMF 1.0),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-RAG-007,rag,poisoning,Can malicious or manipulated documents alter retrieval or generation in unsafe ways?,"Poisoned-document test set, ingestion controls, RAG traces",Poisoning attack success rate,High-impact poisoning scenarios should not cause policy-violating actions or unsupported trusted claims.,critical,pre-deployment,semi-automated,framework-inspired,OWASP Top 10 for LLM Applications 2025,https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-RAG-008,rag,context_overflow,Does the system preserve important evidence when retrieved context exceeds the model context budget?,"Long-context scenarios, selected chunks, outputs",Critical-evidence retention rate,Define critical-evidence retention expectations for long or crowded contexts.,high,pre-deployment,automated,framework-inspired,NIST AI RMF: Generative AI Profile (NIST AI 600-1),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-RAG-009,rag,retrieval_failure,Does the system fail safely when no relevant evidence is available?,"No-answer scenarios, outputs, fallback logs",Appropriate abstention rate,Require abstention or explicit uncertainty where evidence is insufficient for high-impact claims.,high,pre-deployment,automated,framework-inspired,NIST AI RMF: Generative AI Profile (NIST AI 600-1),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-RAG-010,rag,chunking_indexing,"Are chunking, embedding and indexing choices validated for the target corpus and query types?","Ablation tests, retrieval benchmarks, index configuration",Retrieval quality delta,Compare configurations on representative queries; document chosen trade-offs.,medium,pre-deployment,automated,framework-inspired,Hugging Face Evaluate,https://huggingface.co/docs/evaluate/index,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-TOOL-001,tool,schema_validation,Are tool arguments validated against an explicit schema before execution?,"Tool schema, invalid-input tests, execution logs",Invalid-call rejection rate,100% of schema-invalid privileged calls should be rejected before execution.,critical,pre-deployment,automated,framework-inspired,OWASP Top 10 for LLM Applications 2025,https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-TOOL-002,tool,output_validation,Are tool outputs treated as untrusted input and validated before downstream use?,"Output validation rules, malformed-output tests, traces",Unsafe output propagation rate,Zero propagation of known-invalid structured outputs into privileged actions.,critical,pre-deployment,automated,framework-inspired,OWASP Top 10 for LLM Applications 2025,https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-TOOL-003,tool,side_effects,Are side-effecting tools clearly distinguished from read-only tools?,"Tool registry, capability labels, permission rules",Capability labeling coverage,100% of production tools should declare whether they can create external side effects.,high,design,manual,framework-inspired,NIST AI Risk Management Framework (AI RMF 1.0),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-TOOL-004,tool,idempotency,Are retry behaviors safe for tools that may create duplicate or irreversible effects?,"Retry policy, idempotency keys, fault-injection tests",Duplicate side-effect rate,Zero duplicate critical transactions in retry/failure scenarios.,high,pre-deployment,automated,framework-inspired,NIST AI Risk Management Framework (AI RMF 1.0),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-TOOL-005,tool,authentication,"Are tool credentials and service identities scoped, rotated and protected?","Credential inventory, secret-management config, access logs",Credential policy coverage,100% of production tool integrations should use approved credential management.,critical,deployment,semi-automated,framework-inspired,NIST SP 800-218A — Secure Software Development Practices for Generative AI,https://www.nist.gov/publications/secure-software-development-practices-generative-ai-and-dual-use-foundation-models-ssdf,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-TOOL-006,tool,rate_limits,Are execution frequency and resource limits defined for costly or sensitive tools?,"Rate-limit config, stress tests, alerts",Rate-limit enforcement rate,Zero successful bypasses of configured hard limits in the validation suite.,high,pre-deployment,automated,framework-inspired,OWASP Top 10 for LLM Applications 2025,https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-TOOL-007,tool,error_handling,Are tool errors explicit enough for the agent to avoid unsafe assumptions?,"Error taxonomy, simulated failures, traces",Unsafe continuation after error,"High-impact tool failures should trigger safe fallback, stop or escalation behavior.",high,pre-deployment,automated,framework-inspired,NIST AI Risk Management Framework (AI RMF 1.0),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-TOOL-008,tool,auditability,"Are tool calls logged with identity, arguments, result status and timestamp?","Execution logs, trace IDs, audit schema",Audit log completeness,100% of high-impact tool executions should have complete audit records.,high,deployment,automated,framework-inspired,NIST AI Risk Management Framework (AI RMF 1.0),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-DATA-001,data,provenance,"Can training, fine-tuning and evaluation data be traced to known sources and processing steps?","Dataset card, manifests, lineage records, processing logs",Provenance coverage,Define required lineage fields and target complete coverage for production-critical datasets.,high,design,semi-automated,framework-inspired,NIST AI Risk Management Framework (AI RMF 1.0),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-DATA-002,data,quality,"Have missing values, duplicates, corrupt records and label errors been measured?","Data-quality report, validation scripts, sample review",Error rate by defect type,Set dataset-specific quality thresholds and investigate outliers before use.,high,pre-deployment,automated,framework-inspired,Hugging Face Evaluate,https://huggingface.co/docs/evaluate/index,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-DATA-003,data,representativeness,"Does the dataset cover the populations, domains and conditions relevant to intended use?","Coverage analysis, slice statistics, domain comparison",Coverage by target slice,Document known gaps and avoid claims for materially underrepresented conditions.,high,pre-deployment,semi-automated,framework-inspired,NIST AI Risk Management Framework (AI RMF 1.0),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-DATA-004,data,leakage,Has overlap between training data and evaluation/test data been checked where it could bias results?,"Deduplication/overlap analysis, split manifest",Train-test overlap rate,Define a maximum allowed overlap; benchmark contamination should be investigated and reported.,high,pre-deployment,automated,framework-inspired,NIST AI RMF: Generative AI Profile (NIST AI 600-1),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-DATA-005,data,sensitive_information,Is sensitive or restricted information identified and handled according to policy?,"Data classification, DLP scan, access controls",Unapproved sensitive-record rate,Zero knowingly unapproved sensitive records in datasets not authorized to contain them.,critical,pre-deployment,semi-automated,framework-inspired,OWASP Top 10 for LLM Applications 2025,https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-DATA-006,data,poisoning_integrity,Are ingestion and update pipelines protected against unauthorized or malicious data changes?,"Access controls, hashes, provenance, anomaly checks",Unauthorized modification rate,Zero unauthorized modifications to approved production data sources.,critical,deployment,semi-automated,framework-inspired,OWASP Top 10 for LLM Applications 2025,https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-DATA-007,data,label_consistency,"Where labels or human judgments are used, has agreement and consistency been measured?","Annotation guidelines, multiple annotations, adjudication logs",Inter-annotator agreement,Set an agreement target appropriate to the task and document subjective/ambiguous categories.,medium,pre-deployment,automated,framework-inspired,Hugging Face — Choosing a metric,https://huggingface.co/docs/evaluate/en/choosing_a_metric,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-DATA-008,data,drift,Are changes in production input distributions monitored against the validated data profile?,"Reference distribution, production telemetry, drift report",Population/data drift metric,Define alert thresholds by feature or embedding distribution and revalidate after material drift.,high,monitoring,automated,framework-inspired,NIST AI Risk Management Framework (AI RMF 1.0),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-MM-001,multimodal,cross_modal_consistency,"Do text, image, audio or video outputs remain consistent when modalities describe the same underlying fact?","Paired multimodal test set, consistency labels",Cross-modal consistency rate,Set a minimum consistency target on representative paired examples.,high,pre-deployment,semi-automated,framework-inspired,NIST AI RMF: Generative AI Profile (NIST AI 600-1),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-MM-002,multimodal,image_grounding,"When answering about an image, are claims supported by visible content rather than unsupported assumptions?","Images, questions, region/claim annotations, outputs",Visual grounding error rate,Define a maximum unsupported visual-claim rate for the intended task.,high,pre-deployment,semi-automated,framework-inspired,NIST AI RMF: Generative AI Profile (NIST AI 600-1),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-MM-003,multimodal,audio_transcription,"If speech recognition is used, is transcription quality measured across relevant accents, noise and domains?","Audio benchmark, transcripts, slice labels",WER / slice WER,Set use-case-specific WER targets and examine worst-performing slices.,high,pre-deployment,automated,framework-inspired,Hugging Face — Choosing a metric,https://huggingface.co/docs/evaluate/en/choosing_a_metric,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-MM-004,multimodal,modality_injection,"Can malicious instructions embedded in images, documents or audio override system policies?","Adversarial multimodal inputs, traces, outputs",Cross-modal injection success rate,Privileged policy compromise should be zero in the validation suite.,critical,pre-deployment,semi-automated,framework-inspired,OWASP Top 10 for LLM Applications 2025,https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-MM-005,multimodal,missing_modality,"Does the system fail safely when a required modality is missing, unreadable or corrupted?","Corrupted/missing input scenarios, outputs",Safe failure rate,"All required-modality failures should produce explicit error, abstention or fallback behavior.",high,pre-deployment,automated,framework-inspired,NIST AI Risk Management Framework (AI RMF 1.0),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-MM-006,multimodal,format_robustness,"Is behavior stable across common encoding, resolution, compression or sampling variations?","Perturbed media set, outputs",Performance delta by transformation,Define acceptable degradation for supported media transformations.,medium,pre-deployment,automated,framework-inspired,NIST AI Risk Management Framework (AI RMF 1.0),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-MM-007,multimodal,synthetic_content,Are workflows that depend on content authenticity explicit about limitations of detecting or interpreting synthetic media?,"Product documentation, provenance signals, evaluation set",Detection/attribution performance,Do not make unsupported authenticity guarantees; validate claims on representative data.,high,pre-deployment,semi-automated,framework-inspired,NIST AI RMF: Generative AI Profile (NIST AI 600-1),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-MM-008,multimodal,privacy,"Are faces, voices, documents or other sensitive media handled according to access and retention policy?","Data-flow map, retention policy, access logs",Policy violation rate,Zero processing or retention outside approved policy for sensitive multimodal data.,critical,deployment,semi-automated,framework-inspired,NIST AI Risk Management Framework (AI RMF 1.0),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-SYS-001,system,intended_use,"Is the intended use, user population and deployment context explicitly defined?","System card, requirements, use-case documentation",Scope completeness,Define intended use and material out-of-scope uses before validation begins.,high,design,manual,framework-inspired,NIST AI Risk Management Framework (AI RMF 1.0),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-SYS-002,system,risk_ownership,"Are owners assigned for model, data, security, validation and operational risks?","RACI/ownership matrix, governance records",Ownership coverage,Every high-impact risk and control should have a named accountable owner.,high,design,manual,framework-inspired,ISO/IEC 42001:2023 — AI management systems,https://www.iso.org/standard/42001,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-SYS-003,system,change_management,"Do material changes to models, prompts, tools, data or policies trigger revalidation?","Change policy, release records, revalidation logs",Revalidation trigger coverage,100% of predefined material changes should trigger the required revalidation workflow.,high,deployment,semi-automated,framework-inspired,ISO/IEC 42001:2023 — AI management systems,https://www.iso.org/standard/42001,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-SYS-004,system,incident_response,"Can AI incidents be detected, triaged, investigated and linked to model/system evidence?","Incident playbook, logging, drills, postmortems",Detection-to-triage time / evidence completeness,Define incident SLOs and require complete evidence for high-impact events.,critical,deployment,semi-automated,framework-inspired,NIST AI Risk Management Framework (AI RMF 1.0),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-SYS-005,system,monitoring,Are production metrics linked to the assumptions and thresholds used during validation?,"Validation report, dashboards, alert rules",Monitoring coverage,Monitor every production-critical validation criterion that can materially drift over time.,high,monitoring,automated,framework-inspired,NIST AI Risk Management Framework (AI RMF 1.0),https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.
+VAL-SYS-006,system,validation_documentation,"Is there a reproducible validation record covering scope, methods, datasets, metrics, results, limitations and approval?","Validation report, test artifacts, sign-off record",Documentation completeness,No production approval without a traceable validation record and documented residual risks.,high,pre-deployment,manual,framework-inspired,Hugging Face Evaluate,https://huggingface.co/docs/evaluate/index,Practical validation check authored for this dataset; not a verbatim requirement from the cited source.

validation_checklists.jsonl ADDED Viewed

	@@ -0,0 +1,60 @@

+{"id": "VAL-MODEL-001", "system_type": "model", "validation_area": "task_performance", "check": "Does the model meet the minimum quality required for its intended task?", "expected_evidence": "Task-specific evaluation set, baseline model, evaluation report", "metric": "Primary task metric", "threshold_guidance": "Define a use-case-specific minimum and compare against an approved baseline.", "severity": "high", "lifecycle_stage": "pre-deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "Hugging Face — Choosing a metric", "source_url": "https://huggingface.co/docs/evaluate/en/choosing_a_metric", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-MODEL-002", "system_type": "model", "validation_area": "robustness", "check": "Does model performance remain acceptably stable under realistic prompt or input variation?", "expected_evidence": "Perturbation suite, repeated runs, robustness report", "metric": "Performance delta under perturbation", "threshold_guidance": "Define the maximum acceptable degradation for representative variations.", "severity": "high", "lifecycle_stage": "pre-deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "NIST AI Risk Management Framework (AI RMF 1.0)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-MODEL-003", "system_type": "model", "validation_area": "calibration", "check": "Are confidence signals or uncertainty indicators meaningfully aligned with observed correctness where they are used?", "expected_evidence": "Calibration dataset, confidence outputs, reliability analysis", "metric": "Calibration error / selective accuracy", "threshold_guidance": "Set a calibration target appropriate to the decision context; do not use confidence without validation.", "severity": "medium", "lifecycle_stage": "pre-deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "NIST AI Risk Management Framework (AI RMF 1.0)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-MODEL-004", "system_type": "model", "validation_area": "failure_modes", "check": "Are known high-impact failure modes documented and reproducible in testing?", "expected_evidence": "Failure catalog, test prompts, incident examples, model card", "metric": "Failure reproduction coverage", "threshold_guidance": "All identified high-impact failure modes should have at least one reproducible test.", "severity": "high", "lifecycle_stage": "pre-deployment", "automation_level": "semi-automated", "source_type": "framework-inspired", "source_name": "NIST AI RMF: Generative AI Profile (NIST AI 600-1)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-MODEL-005", "system_type": "model", "validation_area": "bias_fairness", "check": "Has performance been evaluated across relevant groups, languages, domains or other meaningful slices?", "expected_evidence": "Slice definitions, evaluation results, disparity analysis", "metric": "Worst-slice performance / disparity", "threshold_guidance": "Define acceptable disparity based on context, impact and population; document unsupported slices.", "severity": "high", "lifecycle_stage": "pre-deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "NIST AI Risk Management Framework (AI RMF 1.0)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-MODEL-006", "system_type": "model", "validation_area": "safety", "check": "Does the model resist known harmful or disallowed behavior within the intended deployment context?", "expected_evidence": "Safety test suite, red-team results, refusal/behavior logs", "metric": "Safety violation rate", "threshold_guidance": "Set scenario-specific limits; critical prohibited behaviors may require zero tolerated occurrences in the test set.", "severity": "critical", "lifecycle_stage": "pre-deployment", "automation_level": "semi-automated", "source_type": "framework-inspired", "source_name": "NIST AI RMF: Generative AI Profile (NIST AI 600-1)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-MODEL-007", "system_type": "model", "validation_area": "versioning", "check": "Can every deployed model instance be traced to an exact model, revision and configuration?", "expected_evidence": "Model registry record, commit/revision ID, deployment manifest", "metric": "Traceability coverage", "threshold_guidance": "100% of production deployments should map to an exact approved model revision.", "severity": "high", "lifecycle_stage": "deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "NIST SP 800-218A — Secure Software Development Practices for Generative AI", "source_url": "https://www.nist.gov/publications/secure-software-development-practices-generative-ai-and-dual-use-foundation-models-ssdf", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-MODEL-008", "system_type": "model", "validation_area": "regression", "check": "Are model updates regression-tested against previously approved capabilities and risks?", "expected_evidence": "Regression suite, prior baseline, change log", "metric": "Regression pass rate / metric deltas", "threshold_guidance": "No release should proceed with unexplained high-impact regressions.", "severity": "high", "lifecycle_stage": "pre-deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "NIST AI Risk Management Framework (AI RMF 1.0)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-MODEL-009", "system_type": "model", "validation_area": "resource_behavior", "check": "Are latency, throughput and memory behavior measured under representative load?", "expected_evidence": "Load-test results, hardware profile, serving logs", "metric": "p95 latency / throughput / memory", "threshold_guidance": "Define operational SLOs for the intended environment and validate under expected concurrency.", "severity": "medium", "lifecycle_stage": "pre-deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "NIST AI Risk Management Framework (AI RMF 1.0)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-MODEL-010", "system_type": "model", "validation_area": "documentation", "check": "Are intended use, limitations, evaluation scope and unsupported conditions documented?", "expected_evidence": "Model card, validation report, risk documentation", "metric": "Documentation completeness", "threshold_guidance": "All deployment-critical limitations and validation boundaries should be documented before release.", "severity": "high", "lifecycle_stage": "pre-deployment", "automation_level": "manual", "source_type": "framework-inspired", "source_name": "NIST AI Risk Management Framework (AI RMF 1.0)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-AGENT-001", "system_type": "agent", "validation_area": "goal_completion", "check": "Does the agent complete representative end-to-end tasks successfully?", "expected_evidence": "Scenario suite, traces, final outcomes", "metric": "Task success rate", "threshold_guidance": "Set a minimum task success rate per workflow and risk level.", "severity": "high", "lifecycle_stage": "pre-deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "NIST AI Risk Management Framework (AI RMF 1.0)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-AGENT-002", "system_type": "agent", "validation_area": "tool_selection", "check": "Does the agent select only appropriate and permitted tools for each task?", "expected_evidence": "Tool-call traces, policy definitions, scenario tests", "metric": "Tool selection accuracy / policy violation rate", "threshold_guidance": "Unauthorized or explicitly disallowed tool selection should be zero in the validation suite.", "severity": "critical", "lifecycle_stage": "pre-deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "OWASP Top 10 for LLM Applications 2025", "source_url": "https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-AGENT-003", "system_type": "agent", "validation_area": "permissions", "check": "Are agent actions constrained by least-privilege permissions?", "expected_evidence": "Permission matrix, service-account scopes, execution logs", "metric": "Unauthorized action rate", "threshold_guidance": "Zero successful actions outside the agent's approved permission scope.", "severity": "critical", "lifecycle_stage": "pre-deployment", "automation_level": "semi-automated", "source_type": "framework-inspired", "source_name": "OWASP Top 10 for LLM Applications 2025", "source_url": "https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-AGENT-004", "system_type": "agent", "validation_area": "trajectory", "check": "Can important agent decisions, tool calls and observations be reconstructed from traces?", "expected_evidence": "Structured traces, tool logs, timestamps, model/version metadata", "metric": "Trace completeness", "threshold_guidance": "100% of high-impact actions should have reconstructable decision and execution evidence.", "severity": "high", "lifecycle_stage": "deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "NIST AI Risk Management Framework (AI RMF 1.0)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-AGENT-005", "system_type": "agent", "validation_area": "recovery", "check": "Does the agent recover safely from tool failures, unavailable dependencies and malformed results?", "expected_evidence": "Fault-injection scenarios, traces, recovery outcomes", "metric": "Safe recovery rate", "threshold_guidance": "Define acceptable recovery behavior; no unsafe fallback actions in critical scenarios.", "severity": "high", "lifecycle_stage": "pre-deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "NIST AI Risk Management Framework (AI RMF 1.0)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-AGENT-006", "system_type": "agent", "validation_area": "termination", "check": "Does the agent stop when its goal is reached or when configured limits are exceeded?", "expected_evidence": "Loop tests, max-step settings, execution logs", "metric": "Unbounded-loop rate / excess-step rate", "threshold_guidance": "Zero unbounded executions; enforce hard runtime, step or cost limits.", "severity": "high", "lifecycle_stage": "pre-deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "OWASP Top 10 for LLM Applications 2025", "source_url": "https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-AGENT-007", "system_type": "agent", "validation_area": "human_oversight", "check": "Are high-impact actions escalated for human approval where required?", "expected_evidence": "Approval policy, approval logs, action traces", "metric": "Approval bypass rate", "threshold_guidance": "Zero execution of actions that policy marks as requiring approval without recorded approval.", "severity": "critical", "lifecycle_stage": "deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "NIST AI Risk Management Framework (AI RMF 1.0)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-AGENT-008", "system_type": "agent", "validation_area": "prompt_injection", "check": "Does the agent resist instructions in untrusted content that attempt to override system or tool-use policies?", "expected_evidence": "Injection test corpus, retrieval/tool scenarios, traces", "metric": "Injection success rate", "threshold_guidance": "Set a strict threshold based on impact; privileged action compromise should be zero in the test suite.", "severity": "critical", "lifecycle_stage": "pre-deployment", "automation_level": "semi-automated", "source_type": "framework-inspired", "source_name": "OWASP Top 10 for LLM Applications 2025", "source_url": "https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-AGENT-009", "system_type": "agent", "validation_area": "cost_control", "check": "Are token, tool, compute and external-service costs bounded and observable?", "expected_evidence": "Budgets, cost telemetry, run logs", "metric": "Cost per successful task / budget breach rate", "threshold_guidance": "Define workflow budgets and require zero uncontrolled budget overruns.", "severity": "medium", "lifecycle_stage": "deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "OWASP Top 10 for LLM Applications 2025", "source_url": "https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-AGENT-010", "system_type": "agent", "validation_area": "delegation", "check": "In multi-agent systems, are delegation boundaries and responsibility transfers explicit and traceable?", "expected_evidence": "Agent graph, delegation events, scopes, traces", "metric": "Delegation trace coverage", "threshold_guidance": "All inter-agent delegation affecting high-impact actions should be attributable and reconstructable.", "severity": "high", "lifecycle_stage": "pre-deployment", "automation_level": "semi-automated", "source_type": "framework-inspired", "source_name": "NIST AI Risk Management Framework (AI RMF 1.0)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-RAG-001", "system_type": "rag", "validation_area": "retrieval_relevance", "check": "Does retrieval return information relevant to the user question or task?", "expected_evidence": "Query set, retrieved passages, relevance judgments", "metric": "Recall@k / nDCG / precision@k", "threshold_guidance": "Choose a retrieval metric and target based on corpus size and downstream task.", "severity": "high", "lifecycle_stage": "pre-deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "Hugging Face — Choosing a metric", "source_url": "https://huggingface.co/docs/evaluate/en/choosing_a_metric", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-RAG-002", "system_type": "rag", "validation_area": "grounding", "check": "Are generated factual claims supported by retrieved evidence when grounding is required?", "expected_evidence": "Answers, retrieved context, claim-evidence labels", "metric": "Grounding error rate / supported-claim rate", "threshold_guidance": "Define a maximum unsupported-claim rate; stricter thresholds for high-impact use cases.", "severity": "critical", "lifecycle_stage": "pre-deployment", "automation_level": "semi-automated", "source_type": "framework-inspired", "source_name": "NIST AI RMF: Generative AI Profile (NIST AI 600-1)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-RAG-003", "system_type": "rag", "validation_area": "citation_accuracy", "check": "Do citations point to sources that actually support the associated claims?", "expected_evidence": "Generated citations, source documents, citation labels", "metric": "Citation precision", "threshold_guidance": "Set a minimum citation-support rate; critical claims should require directly supporting evidence.", "severity": "high", "lifecycle_stage": "pre-deployment", "automation_level": "semi-automated", "source_type": "framework-inspired", "source_name": "NIST AI RMF: Generative AI Profile (NIST AI 600-1)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-RAG-004", "system_type": "rag", "validation_area": "source_provenance", "check": "Can each indexed document be traced to an approved source and version?", "expected_evidence": "Corpus manifest, source URLs, document hashes, ingestion logs", "metric": "Provenance coverage", "threshold_guidance": "100% of production documents should have source and version metadata.", "severity": "high", "lifecycle_stage": "deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "NIST AI Risk Management Framework (AI RMF 1.0)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-RAG-005", "system_type": "rag", "validation_area": "access_control", "check": "Does retrieval enforce document-level access restrictions for the requesting user or agent?", "expected_evidence": "ACLs, test identities, retrieval logs", "metric": "Unauthorized retrieval rate", "threshold_guidance": "Zero retrieval of documents outside the requester's approved access scope.", "severity": "critical", "lifecycle_stage": "pre-deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "OWASP Top 10 for LLM Applications 2025", "source_url": "https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-RAG-006", "system_type": "rag", "validation_area": "staleness", "check": "Are time-sensitive documents refreshed or retired according to defined freshness rules?", "expected_evidence": "Document timestamps, refresh policy, index audit", "metric": "Stale-document rate", "threshold_guidance": "Define freshness windows per source class and alert on expired content.", "severity": "medium", "lifecycle_stage": "monitoring", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "NIST AI Risk Management Framework (AI RMF 1.0)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-RAG-007", "system_type": "rag", "validation_area": "poisoning", "check": "Can malicious or manipulated documents alter retrieval or generation in unsafe ways?", "expected_evidence": "Poisoned-document test set, ingestion controls, RAG traces", "metric": "Poisoning attack success rate", "threshold_guidance": "High-impact poisoning scenarios should not cause policy-violating actions or unsupported trusted claims.", "severity": "critical", "lifecycle_stage": "pre-deployment", "automation_level": "semi-automated", "source_type": "framework-inspired", "source_name": "OWASP Top 10 for LLM Applications 2025", "source_url": "https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-RAG-008", "system_type": "rag", "validation_area": "context_overflow", "check": "Does the system preserve important evidence when retrieved context exceeds the model context budget?", "expected_evidence": "Long-context scenarios, selected chunks, outputs", "metric": "Critical-evidence retention rate", "threshold_guidance": "Define critical-evidence retention expectations for long or crowded contexts.", "severity": "high", "lifecycle_stage": "pre-deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "NIST AI RMF: Generative AI Profile (NIST AI 600-1)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-RAG-009", "system_type": "rag", "validation_area": "retrieval_failure", "check": "Does the system fail safely when no relevant evidence is available?", "expected_evidence": "No-answer scenarios, outputs, fallback logs", "metric": "Appropriate abstention rate", "threshold_guidance": "Require abstention or explicit uncertainty where evidence is insufficient for high-impact claims.", "severity": "high", "lifecycle_stage": "pre-deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "NIST AI RMF: Generative AI Profile (NIST AI 600-1)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-RAG-010", "system_type": "rag", "validation_area": "chunking_indexing", "check": "Are chunking, embedding and indexing choices validated for the target corpus and query types?", "expected_evidence": "Ablation tests, retrieval benchmarks, index configuration", "metric": "Retrieval quality delta", "threshold_guidance": "Compare configurations on representative queries; document chosen trade-offs.", "severity": "medium", "lifecycle_stage": "pre-deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "Hugging Face Evaluate", "source_url": "https://huggingface.co/docs/evaluate/index", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-TOOL-001", "system_type": "tool", "validation_area": "schema_validation", "check": "Are tool arguments validated against an explicit schema before execution?", "expected_evidence": "Tool schema, invalid-input tests, execution logs", "metric": "Invalid-call rejection rate", "threshold_guidance": "100% of schema-invalid privileged calls should be rejected before execution.", "severity": "critical", "lifecycle_stage": "pre-deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "OWASP Top 10 for LLM Applications 2025", "source_url": "https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-TOOL-002", "system_type": "tool", "validation_area": "output_validation", "check": "Are tool outputs treated as untrusted input and validated before downstream use?", "expected_evidence": "Output validation rules, malformed-output tests, traces", "metric": "Unsafe output propagation rate", "threshold_guidance": "Zero propagation of known-invalid structured outputs into privileged actions.", "severity": "critical", "lifecycle_stage": "pre-deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "OWASP Top 10 for LLM Applications 2025", "source_url": "https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-TOOL-003", "system_type": "tool", "validation_area": "side_effects", "check": "Are side-effecting tools clearly distinguished from read-only tools?", "expected_evidence": "Tool registry, capability labels, permission rules", "metric": "Capability labeling coverage", "threshold_guidance": "100% of production tools should declare whether they can create external side effects.", "severity": "high", "lifecycle_stage": "design", "automation_level": "manual", "source_type": "framework-inspired", "source_name": "NIST AI Risk Management Framework (AI RMF 1.0)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-TOOL-004", "system_type": "tool", "validation_area": "idempotency", "check": "Are retry behaviors safe for tools that may create duplicate or irreversible effects?", "expected_evidence": "Retry policy, idempotency keys, fault-injection tests", "metric": "Duplicate side-effect rate", "threshold_guidance": "Zero duplicate critical transactions in retry/failure scenarios.", "severity": "high", "lifecycle_stage": "pre-deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "NIST AI Risk Management Framework (AI RMF 1.0)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-TOOL-005", "system_type": "tool", "validation_area": "authentication", "check": "Are tool credentials and service identities scoped, rotated and protected?", "expected_evidence": "Credential inventory, secret-management config, access logs", "metric": "Credential policy coverage", "threshold_guidance": "100% of production tool integrations should use approved credential management.", "severity": "critical", "lifecycle_stage": "deployment", "automation_level": "semi-automated", "source_type": "framework-inspired", "source_name": "NIST SP 800-218A — Secure Software Development Practices for Generative AI", "source_url": "https://www.nist.gov/publications/secure-software-development-practices-generative-ai-and-dual-use-foundation-models-ssdf", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-TOOL-006", "system_type": "tool", "validation_area": "rate_limits", "check": "Are execution frequency and resource limits defined for costly or sensitive tools?", "expected_evidence": "Rate-limit config, stress tests, alerts", "metric": "Rate-limit enforcement rate", "threshold_guidance": "Zero successful bypasses of configured hard limits in the validation suite.", "severity": "high", "lifecycle_stage": "pre-deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "OWASP Top 10 for LLM Applications 2025", "source_url": "https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-TOOL-007", "system_type": "tool", "validation_area": "error_handling", "check": "Are tool errors explicit enough for the agent to avoid unsafe assumptions?", "expected_evidence": "Error taxonomy, simulated failures, traces", "metric": "Unsafe continuation after error", "threshold_guidance": "High-impact tool failures should trigger safe fallback, stop or escalation behavior.", "severity": "high", "lifecycle_stage": "pre-deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "NIST AI Risk Management Framework (AI RMF 1.0)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-TOOL-008", "system_type": "tool", "validation_area": "auditability", "check": "Are tool calls logged with identity, arguments, result status and timestamp?", "expected_evidence": "Execution logs, trace IDs, audit schema", "metric": "Audit log completeness", "threshold_guidance": "100% of high-impact tool executions should have complete audit records.", "severity": "high", "lifecycle_stage": "deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "NIST AI Risk Management Framework (AI RMF 1.0)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-DATA-001", "system_type": "data", "validation_area": "provenance", "check": "Can training, fine-tuning and evaluation data be traced to known sources and processing steps?", "expected_evidence": "Dataset card, manifests, lineage records, processing logs", "metric": "Provenance coverage", "threshold_guidance": "Define required lineage fields and target complete coverage for production-critical datasets.", "severity": "high", "lifecycle_stage": "design", "automation_level": "semi-automated", "source_type": "framework-inspired", "source_name": "NIST AI Risk Management Framework (AI RMF 1.0)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-DATA-002", "system_type": "data", "validation_area": "quality", "check": "Have missing values, duplicates, corrupt records and label errors been measured?", "expected_evidence": "Data-quality report, validation scripts, sample review", "metric": "Error rate by defect type", "threshold_guidance": "Set dataset-specific quality thresholds and investigate outliers before use.", "severity": "high", "lifecycle_stage": "pre-deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "Hugging Face Evaluate", "source_url": "https://huggingface.co/docs/evaluate/index", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-DATA-003", "system_type": "data", "validation_area": "representativeness", "check": "Does the dataset cover the populations, domains and conditions relevant to intended use?", "expected_evidence": "Coverage analysis, slice statistics, domain comparison", "metric": "Coverage by target slice", "threshold_guidance": "Document known gaps and avoid claims for materially underrepresented conditions.", "severity": "high", "lifecycle_stage": "pre-deployment", "automation_level": "semi-automated", "source_type": "framework-inspired", "source_name": "NIST AI Risk Management Framework (AI RMF 1.0)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-DATA-004", "system_type": "data", "validation_area": "leakage", "check": "Has overlap between training data and evaluation/test data been checked where it could bias results?", "expected_evidence": "Deduplication/overlap analysis, split manifest", "metric": "Train-test overlap rate", "threshold_guidance": "Define a maximum allowed overlap; benchmark contamination should be investigated and reported.", "severity": "high", "lifecycle_stage": "pre-deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "NIST AI RMF: Generative AI Profile (NIST AI 600-1)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-DATA-005", "system_type": "data", "validation_area": "sensitive_information", "check": "Is sensitive or restricted information identified and handled according to policy?", "expected_evidence": "Data classification, DLP scan, access controls", "metric": "Unapproved sensitive-record rate", "threshold_guidance": "Zero knowingly unapproved sensitive records in datasets not authorized to contain them.", "severity": "critical", "lifecycle_stage": "pre-deployment", "automation_level": "semi-automated", "source_type": "framework-inspired", "source_name": "OWASP Top 10 for LLM Applications 2025", "source_url": "https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-DATA-006", "system_type": "data", "validation_area": "poisoning_integrity", "check": "Are ingestion and update pipelines protected against unauthorized or malicious data changes?", "expected_evidence": "Access controls, hashes, provenance, anomaly checks", "metric": "Unauthorized modification rate", "threshold_guidance": "Zero unauthorized modifications to approved production data sources.", "severity": "critical", "lifecycle_stage": "deployment", "automation_level": "semi-automated", "source_type": "framework-inspired", "source_name": "OWASP Top 10 for LLM Applications 2025", "source_url": "https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-DATA-007", "system_type": "data", "validation_area": "label_consistency", "check": "Where labels or human judgments are used, has agreement and consistency been measured?", "expected_evidence": "Annotation guidelines, multiple annotations, adjudication logs", "metric": "Inter-annotator agreement", "threshold_guidance": "Set an agreement target appropriate to the task and document subjective/ambiguous categories.", "severity": "medium", "lifecycle_stage": "pre-deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "Hugging Face — Choosing a metric", "source_url": "https://huggingface.co/docs/evaluate/en/choosing_a_metric", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-DATA-008", "system_type": "data", "validation_area": "drift", "check": "Are changes in production input distributions monitored against the validated data profile?", "expected_evidence": "Reference distribution, production telemetry, drift report", "metric": "Population/data drift metric", "threshold_guidance": "Define alert thresholds by feature or embedding distribution and revalidate after material drift.", "severity": "high", "lifecycle_stage": "monitoring", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "NIST AI Risk Management Framework (AI RMF 1.0)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-MM-001", "system_type": "multimodal", "validation_area": "cross_modal_consistency", "check": "Do text, image, audio or video outputs remain consistent when modalities describe the same underlying fact?", "expected_evidence": "Paired multimodal test set, consistency labels", "metric": "Cross-modal consistency rate", "threshold_guidance": "Set a minimum consistency target on representative paired examples.", "severity": "high", "lifecycle_stage": "pre-deployment", "automation_level": "semi-automated", "source_type": "framework-inspired", "source_name": "NIST AI RMF: Generative AI Profile (NIST AI 600-1)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-MM-002", "system_type": "multimodal", "validation_area": "image_grounding", "check": "When answering about an image, are claims supported by visible content rather than unsupported assumptions?", "expected_evidence": "Images, questions, region/claim annotations, outputs", "metric": "Visual grounding error rate", "threshold_guidance": "Define a maximum unsupported visual-claim rate for the intended task.", "severity": "high", "lifecycle_stage": "pre-deployment", "automation_level": "semi-automated", "source_type": "framework-inspired", "source_name": "NIST AI RMF: Generative AI Profile (NIST AI 600-1)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-MM-003", "system_type": "multimodal", "validation_area": "audio_transcription", "check": "If speech recognition is used, is transcription quality measured across relevant accents, noise and domains?", "expected_evidence": "Audio benchmark, transcripts, slice labels", "metric": "WER / slice WER", "threshold_guidance": "Set use-case-specific WER targets and examine worst-performing slices.", "severity": "high", "lifecycle_stage": "pre-deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "Hugging Face — Choosing a metric", "source_url": "https://huggingface.co/docs/evaluate/en/choosing_a_metric", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-MM-004", "system_type": "multimodal", "validation_area": "modality_injection", "check": "Can malicious instructions embedded in images, documents or audio override system policies?", "expected_evidence": "Adversarial multimodal inputs, traces, outputs", "metric": "Cross-modal injection success rate", "threshold_guidance": "Privileged policy compromise should be zero in the validation suite.", "severity": "critical", "lifecycle_stage": "pre-deployment", "automation_level": "semi-automated", "source_type": "framework-inspired", "source_name": "OWASP Top 10 for LLM Applications 2025", "source_url": "https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-MM-005", "system_type": "multimodal", "validation_area": "missing_modality", "check": "Does the system fail safely when a required modality is missing, unreadable or corrupted?", "expected_evidence": "Corrupted/missing input scenarios, outputs", "metric": "Safe failure rate", "threshold_guidance": "All required-modality failures should produce explicit error, abstention or fallback behavior.", "severity": "high", "lifecycle_stage": "pre-deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "NIST AI Risk Management Framework (AI RMF 1.0)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-MM-006", "system_type": "multimodal", "validation_area": "format_robustness", "check": "Is behavior stable across common encoding, resolution, compression or sampling variations?", "expected_evidence": "Perturbed media set, outputs", "metric": "Performance delta by transformation", "threshold_guidance": "Define acceptable degradation for supported media transformations.", "severity": "medium", "lifecycle_stage": "pre-deployment", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "NIST AI Risk Management Framework (AI RMF 1.0)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-MM-007", "system_type": "multimodal", "validation_area": "synthetic_content", "check": "Are workflows that depend on content authenticity explicit about limitations of detecting or interpreting synthetic media?", "expected_evidence": "Product documentation, provenance signals, evaluation set", "metric": "Detection/attribution performance", "threshold_guidance": "Do not make unsupported authenticity guarantees; validate claims on representative data.", "severity": "high", "lifecycle_stage": "pre-deployment", "automation_level": "semi-automated", "source_type": "framework-inspired", "source_name": "NIST AI RMF: Generative AI Profile (NIST AI 600-1)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-MM-008", "system_type": "multimodal", "validation_area": "privacy", "check": "Are faces, voices, documents or other sensitive media handled according to access and retention policy?", "expected_evidence": "Data-flow map, retention policy, access logs", "metric": "Policy violation rate", "threshold_guidance": "Zero processing or retention outside approved policy for sensitive multimodal data.", "severity": "critical", "lifecycle_stage": "deployment", "automation_level": "semi-automated", "source_type": "framework-inspired", "source_name": "NIST AI Risk Management Framework (AI RMF 1.0)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-SYS-001", "system_type": "system", "validation_area": "intended_use", "check": "Is the intended use, user population and deployment context explicitly defined?", "expected_evidence": "System card, requirements, use-case documentation", "metric": "Scope completeness", "threshold_guidance": "Define intended use and material out-of-scope uses before validation begins.", "severity": "high", "lifecycle_stage": "design", "automation_level": "manual", "source_type": "framework-inspired", "source_name": "NIST AI Risk Management Framework (AI RMF 1.0)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-SYS-002", "system_type": "system", "validation_area": "risk_ownership", "check": "Are owners assigned for model, data, security, validation and operational risks?", "expected_evidence": "RACI/ownership matrix, governance records", "metric": "Ownership coverage", "threshold_guidance": "Every high-impact risk and control should have a named accountable owner.", "severity": "high", "lifecycle_stage": "design", "automation_level": "manual", "source_type": "framework-inspired", "source_name": "ISO/IEC 42001:2023 — AI management systems", "source_url": "https://www.iso.org/standard/42001", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-SYS-003", "system_type": "system", "validation_area": "change_management", "check": "Do material changes to models, prompts, tools, data or policies trigger revalidation?", "expected_evidence": "Change policy, release records, revalidation logs", "metric": "Revalidation trigger coverage", "threshold_guidance": "100% of predefined material changes should trigger the required revalidation workflow.", "severity": "high", "lifecycle_stage": "deployment", "automation_level": "semi-automated", "source_type": "framework-inspired", "source_name": "ISO/IEC 42001:2023 — AI management systems", "source_url": "https://www.iso.org/standard/42001", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-SYS-004", "system_type": "system", "validation_area": "incident_response", "check": "Can AI incidents be detected, triaged, investigated and linked to model/system evidence?", "expected_evidence": "Incident playbook, logging, drills, postmortems", "metric": "Detection-to-triage time / evidence completeness", "threshold_guidance": "Define incident SLOs and require complete evidence for high-impact events.", "severity": "critical", "lifecycle_stage": "deployment", "automation_level": "semi-automated", "source_type": "framework-inspired", "source_name": "NIST AI Risk Management Framework (AI RMF 1.0)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-SYS-005", "system_type": "system", "validation_area": "monitoring", "check": "Are production metrics linked to the assumptions and thresholds used during validation?", "expected_evidence": "Validation report, dashboards, alert rules", "metric": "Monitoring coverage", "threshold_guidance": "Monitor every production-critical validation criterion that can materially drift over time.", "severity": "high", "lifecycle_stage": "monitoring", "automation_level": "automated", "source_type": "framework-inspired", "source_name": "NIST AI Risk Management Framework (AI RMF 1.0)", "source_url": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}
+{"id": "VAL-SYS-006", "system_type": "system", "validation_area": "validation_documentation", "check": "Is there a reproducible validation record covering scope, methods, datasets, metrics, results, limitations and approval?", "expected_evidence": "Validation report, test artifacts, sign-off record", "metric": "Documentation completeness", "threshold_guidance": "No production approval without a traceable validation record and documented residual risks.", "severity": "high", "lifecycle_stage": "pre-deployment", "automation_level": "manual", "source_type": "framework-inspired", "source_name": "Hugging Face Evaluate", "source_url": "https://huggingface.co/docs/evaluate/index", "notes": "Practical validation check authored for this dataset; not a verbatim requirement from the cited source."}