AI explainability control library32
HighGOV-01
Explanation policy and accountable ownership
Define which AI-supported decisions require explanations, who must understand them, who approves them, and which laws, policies, and records apply.
HighGOV-02
Explanation requirement by consequence
Set explanation depth, timing, channel, evidence, review, and retention requirements according to consumer and business consequence.
CriticalGOV-03
Purpose limits and knowledge boundaries
State the intended purpose, supported population, assumptions, limitations, uncertainty, and conditions where the system must abstain or escalate.
CriticalGOV-04
Transparency exception and escalation
Document justified limits on disclosure, residual risk, alternative information, approver, expiration, monitoring, and escalation.
HighTRC-01
System and workflow map
Map data, models, prompts, retrieval, rules, tools, people, vendors, decisions, notices, actions, and downstream systems.
HighTRC-02
Input output and version trace
Retain the material inputs, derived values, model and prompt versions, rules, thresholds, outputs, and decision timestamp needed to reconstruct a result.
CriticalTRC-03
Human contribution and authority
Record who reviewed, changed, approved, or rejected AI-supported work and what information was available at the time.
CriticalTRC-04
End-to-end decision reconstruction
Demonstrate that a qualified reviewer can reproduce the sequence from source evidence through AI output, human judgment, action, and notice.
HighPRED-01
Global model behavior
Explain the model purpose, training and validation population, material variables, interactions, thresholds, performance, and limitations.
HighPRED-02
Local prediction reasons
Identify the factors and values that materially influenced an individual score, classification, referral, or recommendation.
CriticalPRED-03
Reason code fidelity
Verify that stated reasons accurately reflect the actual model, rule, threshold, and workflow drivers for the individual outcome.
CriticalPRED-04
Stability sensitivity and uncertainty
Test whether explanations remain reasonably stable under immaterial changes and communicate uncertainty or low confidence appropriately.
HighRAG-01
Source attribution and citation fidelity
Link material generated statements to approved sources and verify that citations support the associated claim.
HighRAG-02
Retrieval context transparency
Record collections, permissions, query, retrieved content, ranking, freshness, filters, and missing-source conditions relevant to the response.
CriticalRAG-03
Generated content boundaries
Clearly distinguish generated text from verified facts, approved language, human edits, and unresolved uncertainty.
CriticalRAG-04
Hallucination and unsupported claim handling
Detect unsupported material claims, abstain or qualify when evidence is insufficient, and route consequential content for review.
HighAGT-01
Goal plan and delegation trace
Record the approved goal, plan steps, memory, delegation, changes in approach, budgets, and termination conditions.
HighAGT-02
Tool call and transaction record
Capture tool identity, target, parameters, authorization, result, side effects, errors, retries, approvals, and transaction identifiers.
CriticalAGT-03
Policy and approval visibility
Show which policy, rule, permission, threshold, and human approval allowed, denied, or modified an action.
CriticalAGT-04
Partial action and recovery explanation
Explain what completed, failed, rolled back, remains pending, and requires reconciliation after an interrupted or failed workflow.
HighCOM-01
Consumer notice and adverse-action reasons
Provide specific, timely, understandable reasons and identify material data sources, correction rights, contact channels, and recourse where required.
HighCOM-02
Insurance professional decision support
Present the evidence, limitations, uncertainty, sources, and next steps an adjuster, underwriter, or reviewer needs to exercise judgment.
CriticalCOM-03
Regulatory and audit documentation
Maintain documentation that lets reviewers understand the system, its use, controls, tests, changes, outcomes, and specific sampled decisions.
CriticalCOM-04
Accessibility language and channel
Test readability, disability access, language, format, delivery, and availability of human assistance for intended audiences.
HighVAL-01
Fidelity and completeness testing
Test whether the explanation reflects the actual process and covers the material factors, sources, rules, and actions behind the outcome.
HighVAL-02
Comprehension and usability testing
Evaluate whether intended users can understand the explanation, identify its limits, and take the expected action.
CriticalVAL-03
Consistency and discrimination review
Compare similar cases and relevant segments for inconsistent, misleading, or different-quality explanations and access to recourse.
CriticalVAL-04
Reviewer automation-bias testing
Assess whether explanation design encourages uncritical acceptance, hides uncertainty, or creates false confidence in AI-supported work.
HighOPS-01
Production explanation monitoring
Track missing, generic, contradictory, unsupported, stale, inaccessible, or disputed explanations and their consumer or operational impact.
HighOPS-02
Change and regression control
Repeat explanation and traceability tests after material changes to models, data, prompts, retrieval, rules, tools, vendors, notices, or workflows.
CriticalOPS-03
Vendor transparency and evidence rights
Require documentation, trace access, explanation support, testing, limitations, change notice, audit cooperation, correction, and exit rights.
CriticalOPS-04
Correction affected-case review and closure
Correct explanations and underlying decisions, identify affected cases, communicate as required, retest controls, and close root causes.