SWT3 Protocol

NIST AI RMF
Assessor Workpaper

AI Risk Management Framework -- Cryptographic Evidence Verification

Engagement Details

Client Organization
Assessment Date(s)
Lead Assessor
Assessment Organization
AI RMF Scope

Workpaper Usage by AI RMF Function

FunctionFocusWorkpaper Entries
GovernPolicy and governance entriesWP-01 through WP-10
MapRisk identification and contextWP-11 through WP-16
MeasureMetrics, testing, and monitoringWP-17 through WP-19
ManageOperations and incident responseWP-20 through WP-25

Entry Cross-References

EntryDepends On / See Also
WP-05 (Guardrails)WP-19 (Red Team) -- discovered vulnerabilities should update guardrail config
WP-10 (Enforcement)WP-25 (Violations) -- violations must flow to enforcement
WP-11 (Inventory)WP-20 (Model Lifecycle) -- every production model must be registered
WP-16 (Impact Assessment)WP-12 (Risk Identification) -- impact assessments require prior risk identification
WP-17 (Performance)WP-18 (Drift) -- drift is measured against performance baselines
WP-18 (Drift)WP-22 (Safety) -- drift threshold exceedances should trigger safety reviews
WP-19 (Red Team)WP-21 (Cybersecurity) -- red team findings inform threat assessments
WP-24 (Incident)WP-09 (Override) -- incidents may require human override
Critical Assessor Notice: Boundaries of Cryptographic Evidence

SWT3 witness anchors prove that specific operational controls were active at a specific point in time. They do not replace the assessor's independent judgment, professional expertise, or regulatory authority. Assessment determination ratings remain the assessor's responsibility.

Version 1.0 | August 2026 | Tenable Nova LLC | sovereign.tenova.io

Instructions

This workpaper provides structured test procedures for verifying SWT3 witness anchor evidence during a NIST AI Risk Management Framework (AI 100-1) assessment. Each entry maps an AI RMF subcategory to the SWT3 procedures that produce evidence, defines what to test, what to expect, and how to determine pass/fail.

How to verify an anchor: Paste the anchor string into the public verifier at sovereign.tenova.io/verify (no account required), or recompute the fingerprint yourself using SHA-256. See Verify an Anchor in 5 Seconds for the full procedure.

Coverage badges: Full = SWT3 procedure directly satisfies the subcategory requirement. Partial = procedure covers part of the subcategory; assessor must verify remaining elements through interview or document review.

Sampling guidance: For sample size, red flags, and CSV query instructions, see Assessor Sampling Instructions.

Factor conventions: Factor legends in this workpaper (e.g., "factor_a = policy version") reflect the SWT3 SDK default conventions. Clients may customize factor usage. Before relying on factor interpretations, confirm with the client that their implementation follows these defaults, or request their factor mapping document.

Methodology boundary: This workpaper documents evidence format and suggests evaluation criteria. It does not prescribe assessment methodology. Sufficiency determinations and assessment conclusions remain the assessor's professional responsibility under their applicable accreditation requirements.

Not applicable entries: If a workpaper entry is not applicable to the assessment scope (e.g., the organization does not operate multi-agent systems, so WP-03 supply chain entries do not apply), mark the result as N/A with justification in the Notes field.

Estimated duration: Allow 3-4 hours for a complete workpaper review including verification of 5-10 sample anchors per function. For initial assessments, plan for an additional 1-2 hours of orientation with the provider's factor mapping and ledger access. Periodic reviews with an established baseline typically require 2-3 hours.

Anchor anatomy: Every SWT3 Witness Anchor follows a deterministic format. Here is an annotated example for an AI RMF-relevant procedure:

SWT3-E-AWS-AI-DRIFT1-PASS-1780300000-a3f8c92b1d07
SegmentMeaning
SWT3-EEnclave deployment tier
AWSInfrastructure provider
AIAI procedures namespace
DRIFT1AI-DRIFT.1 procedure (MEASURE 2.6)
PASSAcceptance criteria met
1780300000Unix epoch timestamp
a3f8c92b1d07First 12 chars of SHA-256 fingerprint

Govern Govern Function

The Govern function establishes the organizational context, policies, and oversight structures for AI risk management. Entries WP-01 through WP-10 address governance policies, technical environment, supply chain, audit, trust, transparency, roles, governance mechanisms, human oversight, and enforcement.

Control Objective
The organization shall establish policies addressing dual-use potential, acceptable use governance, and human-in-the-loop oversight for AI systems.
SWT3 Procedures
AI-DUALUSE.1 -- Dual-Use Classification | AI-GOV.1 -- Acceptable Use Policy Attestation | AI-HITL.1 -- Human-in-the-Loop Oversight
Factor Legend
AI-DUALUSE.1: factor_a = risk tier (1-4) | factor_b = classification date hash | factor_c = justification length
AI-GOV.1: factor_a = policy version | factor_b = compliance status (1 = compliant) | factor_c = days since review
AI-HITL.1: factor_a = reviewer identity hash | factor_b = decision (1=approved, 0=rejected) | factor_c = review duration (seconds) (SDK defaults -- confirm with provider)
Test Procedure
  • Request AI-DUALUSE.1 anchors for systems with dual-use potential and verify that classification predates deployment.
  • Confirm AI-GOV.1 policy is current (factor_b = 1) and review cycle is within organizational policy.
  • Verify AI-HITL.1 anchors match the operational workflow described in the organization's documentation.
  • Independently verify one anchor fingerprint using the public verifier.
Expected Evidence
Dual-use classification anchors predating deployment for relevant systems. Current policy attestation with compliant status. Human oversight records consistent with stated workflow.
Sample Verification
sovereign.tenova.io/verify -- paste any AI-DUALUSE.1, AI-GOV.1, or AI-HITL.1 anchor token
Pass/Fail Criteria
PASS if: Dual-use is classified before deployment AND policy is current AND HITL oversight matches documented workflow. FAIL if: No dual-use classification for systems with dual-use potential, or policy has lapsed without renewal. SWT3 proves the dual-use classification was recorded and the policy is current. The assessor evaluates whether the classification is appropriate for the system's capabilities and deployment context.
Red Flag
Watch for: All AI-GOV.1 factor_b values show compliant but factor_c (days since review) exceeds 365 -- policy may be auto-renewed without actual review.
Notes / Findings
Control Objective
The organization shall document the technical environment, model provenance, and infrastructure dependencies for each AI system.
SWT3 Procedures
AI-ENV.1 -- Environment Attestation | AI-HW.1 -- Hardware Attestation | AI-MDL.5 -- Model Weight Attestation | AI-MDL.6 -- Adapter Stack Attestation
Factor Legend
AI-ENV.1: factor_a = environment identifier | factor_b = configuration hash | factor_c = runtime version
AI-HW.1: factor_a = hardware platform | factor_b = accelerator type | factor_c = firmware version
AI-MDL.5: factor_a = model architecture | factor_b = weight file SHA-256 hash | factor_c = training framework/version
AI-MDL.6: factor_a = adapter name | factor_b = adapter hash | factor_c = base model reference (SDK defaults -- confirm with provider)
Test Procedure
  • Compare AI-MDL.5 weight hash (factor_b) against the hash of the production model file.
  • Verify per-environment anchors exist for each deployment target (dev, staging, production).
  • Check adapter lineage via AI-MDL.6 to confirm fine-tuned models trace back to documented base models.
  • Verify AI-HW.1 reflects the actual hardware in use.
Expected Evidence
Model weight hashes consistent between development and production. Per-environment attestation covering all deployment targets. Adapter lineage tracing to documented base models.
Sample Verification
sovereign.tenova.io/verify -- paste any AI-MDL.5 or AI-ENV.1 anchor token
Pass/Fail Criteria
PASS if: Weight hashes are consistent across environments and all environments are documented. FAIL if: Model changes are untracked, or weight hashes differ between declared and deployed models without justification.
Red Flag
Watch for: AI-MDL.5 weight hashes are identical across months for a model described as 'continuously fine-tuned.' Hashes should change after each training cycle.
Notes / Findings
Control Objective
The organization shall document and attest to the supply chain for AI components, including third-party models, and shall track multi-agent interactions with complete lifecycle chains.
SWT3 Procedures
AI-CHAIN.1 -- Supply Chain Attestation | AI-CHAIN.2 -- Chain Lifecycle | AI-GOV.6 -- Supply Chain Governance | AI-MULTI.1 -- Multi-Agent Coordination
Factor Legend
AI-CHAIN.1: factor_a = supplier identifier | factor_b = attestation status (1=verified) | factor_c = component count
AI-CHAIN.2: factor_a = cycle_id | factor_b = sequence position | factor_c = chain length
AI-MULTI.1: factor_a = agent count | factor_b = orchestration pattern | factor_c = handoff count (SDK defaults -- confirm with provider)
Test Procedure
  • Verify AI-CHAIN.2 cycle_ids form complete sequences (no missing positions in the chain).
  • Check AI-CHAIN.1 supply chain attestations for all third-party AI components.
  • For multi-agent systems, verify AI-MULTI.1 documents all agent handoffs.
  • Confirm AI-GOV.6 governance attestation covers supply chain management policies.
Expected Evidence
Complete lifecycle chains with no gaps in sequence positions. Supply chain attestation for every third-party component. Multi-agent handoffs documented with agent counts.
Sample Verification
sovereign.tenova.io/verify -- paste any AI-CHAIN.1 or AI-CHAIN.2 anchor token
Pass/Fail Criteria
PASS if: Chains are complete and supply chain is attested for all third-party components. FAIL if: Unwitnessed handoffs exist in multi-agent systems, or third-party components lack attestation. SWT3 proves supply chain attestations exist and lifecycle chains are complete. The assessor evaluates whether attested suppliers meet organizational and regulatory requirements.
Red Flag
Watch for: AI-CHAIN.2 chains always show sequence length of 1 in a declared multi-agent system. This suggests only the first agent is witnessed.
Notes / Findings
Control Objective
The organization shall maintain audit trails and logging infrastructure for AI system operations, with integrity verification and external timestamping.
SWT3 Procedures
AI-AUDIT.1 -- Audit Log Integrity | AI-AUDIT.2 -- External Timestamp | AI-LOG.1 -- Logging Pipeline
Factor Legend
AI-AUDIT.1: factor_a = log entry count | factor_b = integrity hash (1=verified intact) | factor_c = retention period (days)
AI-AUDIT.2: factor_a = timestamp authority | factor_b = verification status (1=confirmed) | factor_c = time delta (seconds from trusted source)
AI-LOG.1: factor_a = pipeline identifier | factor_b = events processed | factor_c = drop rate (0=no drops) (SDK defaults -- confirm with provider)
Test Procedure
  • Verify AI-AUDIT.1 anchors confirm audit log integrity on a regular cadence.
  • Check AI-AUDIT.2 for RFC 3161 timestamp verification (factor_a should identify a trusted TSA).
  • Confirm AI-LOG.1 shows the logging pipeline is operational with zero or near-zero drop rate.
  • Verify log retention (factor_c in AI-AUDIT.1) meets organizational and regulatory requirements.
Expected Evidence
Regular audit integrity checks with verified hashes. External timestamp confirmation from a trusted authority. Logging pipeline operational with minimal data loss.
Sample Verification
sovereign.tenova.io/verify -- paste any AI-AUDIT.1 or AI-AUDIT.2 anchor token
Pass/Fail Criteria
PASS if: Audit integrity is verified at the organization's stated frequency AND RFC 3161 timestamps are confirmed AND logging pipeline is operational. FAIL if: No audit evidence exists, or significant data loss is indicated by the drop rate.
Red Flag
Watch for: AI-LOG.1 drop rate suddenly improves from >0 to exactly 0 on a specific date. May indicate the drop counter was reset rather than the issue being fixed.
Notes / Findings
Control Objective
The organization shall enforce guardrails at inference time, implement pre-inference policy gates, and verify trust between AI system components.
SWT3 Procedures
AI-GRD.1 -- Guardrail Attestation | AI-GRD.2 -- Output Filtering | AI-TRUST.1 -- Trust Verification | AI-GOV.4 -- Pre-Inference Policy Gate
Factor Legend
AI-GRD.1: factor_a = guardrail type | factor_b = test result (1=active) | factor_c = blocked content category
AI-GRD.2: factor_a = filter type | factor_b = items filtered | factor_c = false positive rate
AI-TRUST.1: factor_a = verification checks total | factor_b = checks passed | factor_c = trust score
AI-GOV.4: factor_a = policy gate identifier | factor_b = authorization result (1=authorized) | factor_c = authorization_id (SDK defaults -- confirm with provider)
Test Procedure
  • Verify AI-GRD.1 anchors confirm guardrails are active at inference time (factor_b = 1).
  • Check AI-GOV.4 timestamps to confirm policy gate fires before inference anchors.
  • Verify AI-TRUST.1 trust pairs between system components.
  • Confirm AI-GRD.2 output filtering is operational with acceptable false positive rate.
Expected Evidence
Guardrail attestation showing continuous enforcement during inference. Policy gate anchors timestamped before corresponding inference anchors. Trust verification between all connected components.
Sample Verification
sovereign.tenova.io/verify -- paste any AI-GRD.1 or AI-GOV.4 anchor token
Pass/Fail Criteria
PASS if: Guardrails are continuously enforced AND policy gate fires before inference AND trust pairs are verified. FAIL if: Guardrails are undocumented at runtime, or policy gate is absent for high-risk inferences. SWT3 proves guardrails are active and policy gates fire before inference. The assessor evaluates whether the guardrail configuration is appropriate for the system's risk profile.
Red Flag
Watch for: AI-GOV.4 policy gate timestamps are identical to AI-INF.1 inference timestamps to the millisecond. A real gate check takes measurable time.
Notes / Findings
Control Objective
The organization shall maintain transparency through agent charters, content marking, tool documentation, and license compliance for AI systems.
SWT3 Procedures
AI-CHR.1 -- Agent Charter | AI-MARK.1 -- Content Marking | AI-TOOL.1 -- Tool Witnessing | AI-TRANS.1 -- Transparency Attestation | AI-LIC.1 -- License Compliance
Factor Legend
AI-CHR.1: factor_a = charter version | factor_b = scope hash | factor_c = effective date
AI-MARK.1: factor_a = marking type | factor_b = marking applied (1=yes) | factor_c = content category
AI-LIC.1: factor_a = license identifier | factor_b = compliance status (1=compliant) | factor_c = model source (SDK defaults -- confirm with provider)
Test Procedure
  • Verify AI-CHR.1 agent charters predate the first inference anchor for each agent.
  • Check AI-LIC.1 license compliance for open-weight models (factor_b = 1).
  • Confirm AI-MARK.1 content marking is active for AI-generated outputs.
  • Verify AI-TOOL.1 anchors exist for each tool invoked by AI agents.
Expected Evidence
Agent charters predating deployment. License compliance attestation for all models. Content marking active on AI-generated outputs. Tool invocations witnessed. SWT3 proves that documentation exists and content is marked. The assessor evaluates documentation quality, marking adequacy, and whether transparency disclosures meet stakeholder needs.
Sample Verification
sovereign.tenova.io/verify -- paste any AI-CHR.1 or AI-LIC.1 anchor token
Pass/Fail Criteria
PASS if: Documentation anchors exist and content marking is active for AI-generated content. FAIL if: Undocumented agents are in production, or unlicensed model usage is detected.
Red Flag
Watch for: AI-TOOL.1 anchors exist for 3 of 12 declared tools. The 9 unwitnessed tools may be operating without governance.
Notes / Findings
Control Objective
The organization shall define and assign roles and responsibilities for AI risk management, including named individuals accountable for AI system oversight.
SWT3 Procedures
AI-GOV.2 -- Role Assignment Attestation | AI-INF.3 -- Authorized Inference Personnel
Factor Legend
AI-GOV.2: factor_a = role identifier | factor_b = assignee hash | factor_c = scope description hash
AI-INF.3: factor_a = personnel identifier hash | factor_b = authorization level | factor_c = system scope (SDK defaults -- confirm with provider)
Test Procedure
  • Confirm AI-GOV.2 anchors reference named individuals, not generic teams.
  • Verify AI-INF.3 anchors confirm authorized personnel for high-risk inference operations.
  • Cross-reference role assignments with the organization's RACI matrix or equivalent.
Expected Evidence
Role assignment anchors referencing named individuals with documented scope. Authorized inference personnel documented for high-risk operations. SWT3 proves role assignment occurred. The assessor evaluates whether assigned roles have sufficient authority and whether the RACI matrix covers all AI RMF functions.
Sample Verification
sovereign.tenova.io/verify -- paste any AI-GOV.2 or AI-INF.3 anchor token
Pass/Fail Criteria
PASS if: Roles are assigned to named parties with documented scope. FAIL if: Accountability is generic or unnamed for AI risk management responsibilities.
Red Flag
Watch for: AI-GOV.2 role assignments reference the same person hash for all roles. One person holding all AI governance roles defeats segregation of duties.
Notes / Findings
Control Objective
The organization shall implement governance mechanisms, including governance boards or committees, that make documented decisions about AI system lifecycle events.
SWT3 Procedure
AI-GOV.7 -- Governance Board Decision
Factor Legend
AI-GOV.7: factor_a = decision type | factor_b = approval status (1=approved, 0=denied) | factor_c = board quorum hash (SDK defaults -- confirm with provider)
Test Procedure
  • Check AI-GOV.7 anchors for governance board decisions covering major AI lifecycle events (deployment, model change, retirement).
  • Verify cadence via timestamp distribution -- governance decisions should correspond to lifecycle events.
  • Cross-reference with meeting minutes or governance documentation.
Expected Evidence
Governance board decisions documented for major lifecycle events. Decision anchors with approval status and quorum verification. SWT3 proves a governance decision was recorded. The assessor evaluates whether the board has appropriate composition and authority.
Sample Verification
sovereign.tenova.io/verify -- paste any AI-GOV.7 anchor token
Pass/Fail Criteria
PASS if: Governance decisions are documented for major lifecycle events. FAIL if: No governance evidence exists for deployment approvals or major model changes.
Red Flag
Watch for: AI-GOV.7 governance board decisions are always approved (factor_b = 1) with no denials across the entire assessment period.
Notes / Findings
Control Objective
The organization shall implement and test human override mechanisms for AI systems, ensuring that human operators can intervene or shut down AI operations when necessary.
SWT3 Procedure
AI-HITL.3 -- Human Override
Factor Legend
AI-HITL.3: factor_a = override type | factor_b = override authority hash | factor_c = response time (seconds) (SDK defaults -- confirm with provider)
Test Procedure
  • Request AI-HITL.3 override records for the assessment period.
  • Cross-reference override frequency with the system's risk tier -- high-risk systems should show evidence of tested override mechanisms.
  • Note: Zero overrides over extended periods may indicate the mechanism is unavailable rather than unused. Request evidence of periodic override testing if no operational overrides exist.
Expected Evidence
Override records or documented override testing (tabletop/drill) for the assessment period. Response times documented. SWT3 proves the override occurred. The assessor evaluates whether the override mechanism is adequate for the system's risk profile and whether personnel are trained to exercise it.
Sample Verification
sovereign.tenova.io/verify -- paste any AI-HITL.3 anchor token
Pass/Fail Criteria
PASS if: Override mechanism is documented and exercised (or documented as tested via tabletop/drill). FAIL if: No override capability exists for high-risk systems.
Red Flag
Watch for: Zero AI-HITL.3 override records over 6+ months for a high-risk system. The override mechanism may exist in documentation but not in practice.
Notes / Findings
Control Objective
The organization shall enforce AI policies through documented enforcement actions, ensuring that policy violations result in corrective measures.
SWT3 Procedure
AI-GOV.5 -- Policy Enforcement Attestation
Factor Legend
AI-GOV.5: factor_a = enforcement action type | factor_b = action taken (1=enforced) | factor_c = related violation reference (SDK defaults -- confirm with provider)
Test Procedure
  • Verify AI-GOV.5 enforcement actions exist for detected policy violations.
  • Cross-reference with AI-VIO.1 violation records from the MANAGE function (WP-25).
  • Confirm violations flow through to enforcement -- detected violations without corresponding enforcement are a gap.
Expected Evidence
Enforcement action anchors corresponding to detected violations. Violation-to-enforcement traceability. SWT3 proves enforcement actions were recorded. The assessor evaluates whether enforcement actions are proportionate and effective.
Sample Verification
sovereign.tenova.io/verify -- paste any AI-GOV.5 anchor token
Pass/Fail Criteria
PASS if: Violations flow through to documented enforcement actions. FAIL if: Violations are detected but never enforced.
Red Flag
Watch for: AI-GOV.5 enforcement actions exist but AI-VIO.1 violations from WP-25 have no corresponding enforcement records. Violations may be detected but not acted upon.
Notes / Findings

Map Map Function

The Map function identifies and contextualizes AI risks. Entries WP-11 through WP-16 address system identification, risk identification, fairness and explainability, data governance, data lineage, and impact assessment.

Control Objective
The organization shall maintain a complete inventory of AI systems, including persistent identifiers, software bills of materials, and model cards.
SWT3 Procedures
AI-GOV.3 -- System Registration | AI-ID.1 -- Agent Identity | AI-MDL.4 -- Model Card | AI-SBOM.1 -- AI Software Bill of Materials
Factor Legend
AI-ID.1: factor_a = agent_id (persistent identifier) | factor_b = agent type | factor_c = registration timestamp
AI-SBOM.1: factor_a = component count | factor_b = vulnerability count | factor_c = last scan timestamp
AI-MDL.4: factor_a = model name | factor_b = model version | factor_c = card completeness score (SDK defaults -- confirm with provider)
Test Procedure
  • Confirm every AI system in production has a corresponding AI-ID.1 anchor with a persistent identifier.
  • Verify AI-SBOM.1 is current (factor_c indicates recent scan).
  • Check AI-MDL.4 model cards cover all production models.
  • Verify AI-GOV.3 system registration anchors exist for each system.
Expected Evidence
Complete system inventory with persistent identifiers. Current SBOMs for all AI systems. Model cards covering all production models.
Sample Verification
sovereign.tenova.io/verify -- paste any AI-ID.1 or AI-SBOM.1 anchor token
Pass/Fail Criteria
PASS if: All systems are identified with persistent IDs and current SBOMs. FAIL if: Unregistered models are found in production, or SBOMs are stale.
Red Flag
Watch for: AI-SBOM.1 was minted once at deployment and never refreshed. Dependencies change -- a stale SBOM does not reflect the current attack surface.
Notes / Findings
Control Objective
The organization shall identify and categorize AI risks across technical, fairness, privacy, and societal dimensions at deployment and on a recurring schedule.
SWT3 Procedure
AI-RISK.1 -- Risk Identification
Factor Legend
AI-RISK.1: factor_a = risk category | factor_b = risk count identified | factor_c = mitigation coverage (0-1 ratio) (SDK defaults -- confirm with provider)
Test Procedure
  • Confirm AI-RISK.1 anchors exist at deployment and on a recurring schedule.
  • Review risk categories (factor_a) for completeness -- expect technical, fairness, privacy, and societal dimensions.
  • Verify mitigation coverage ratio (factor_c) is tracked and improving over time.
Expected Evidence
Risk identification anchors categorized across multiple risk dimensions. Mitigation coverage ratio documented. Recurring assessment schedule demonstrated through anchor timestamps.
Sample Verification
sovereign.tenova.io/verify -- paste any AI-RISK.1 anchor token
Pass/Fail Criteria
PASS if: Risk assessments cover technical, fairness, privacy, and societal risks with documented mitigation. FAIL if: Assessments cover only technical risks, or no recurring assessment schedule is demonstrated.
Red Flag
Watch for: AI-RISK.1 assessments only cover technical risks (model failure, latency). Fairness, privacy, and societal impact categories are absent.
Notes / Findings
Control Objective
The organization shall test for fairness across protected attributes, document explainability methods, and verify retrieval-augmented generation data provenance.
SWT3 Procedures
AI-FAIR.2 -- Bias Testing | AI-FAIR.3 -- Disparate Impact | AI-EXPL.2 -- Explanation Quality | AI-RAG.1 -- Retrieval Provenance
Factor Legend
AI-FAIR.2: factor_a = protected attribute | factor_b = statistical test | factor_c = threshold
AI-FAIR.3: factor_a = protected group | factor_b = impact ratio | factor_c = four-fifths threshold status
AI-EXPL.2: factor_a = explanation method | factor_b = quality score | factor_c = evaluation criteria
AI-RAG.1: factor_a = retrieval source | factor_b = documents retrieved | factor_c = relevance score (SDK defaults -- confirm with provider)
Test Procedure
  • Verify AI-FAIR.2 fairness testing covers all relevant protected attributes.
  • Check AI-FAIR.3 for disparate impact analysis with documented thresholds.
  • Confirm AI-EXPL.2 documents the explainability method and quality assessment.
  • Verify AI-RAG.1 retrieval provenance for RAG-based systems.
Expected Evidence
Fairness testing across protected attributes with statistical methods and thresholds. Explainability documentation with quality assessment. RAG provenance covering retrieval sources and relevance.
Sample Verification
sovereign.tenova.io/verify -- paste any AI-FAIR.2 or AI-RAG.1 anchor token
Pass/Fail Criteria
PASS if: Fairness and explainability are documented for consequential systems with defined thresholds. FAIL if: No fairness evidence exists for systems affecting individuals, or explainability is absent for high-risk decisions.
Red Flag
Watch for: AI-FAIR.2 tests only one protected attribute (e.g., gender) when the system processes data about multiple protected groups.
Notes / Findings
Control Objective
The organization shall implement data governance for AI systems, including data quality attestation and synthetic data declarations.
SWT3 Procedures
AI-DATA.1 -- Data Provenance | AI-DATA.4 -- Synthetic Data Declaration
Factor Legend
AI-DATA.1: factor_a = data source identifier | factor_b = record count | factor_c = collection method
AI-DATA.4: factor_a = synthetic data ratio | factor_b = generation method | factor_c = quality validation status (SDK defaults -- confirm with provider)
Test Procedure
  • Confirm AI-DATA.1 anchors document data sources, volumes, and collection methods for each dataset used in AI systems.
  • Verify AI-DATA.4 synthetic data declarations identify the ratio, generation method, and quality validation.
  • Cross-reference data governance anchors with the organization's data management policy.
Expected Evidence
Data provenance covering all data sources with collection methods documented. Synthetic data clearly declared with generation methods and quality validation.
Sample Verification
sovereign.tenova.io/verify -- paste any AI-DATA.1 or AI-DATA.4 anchor token
Pass/Fail Criteria
PASS if: Data sources are documented with quality metrics for all active datasets. FAIL if: Training data is undocumented, or synthetic data is used without declaration.
Red Flag
Watch for: AI-DATA.4 synthetic data declaration shows 0% synthetic but the training dataset is suspiciously large for the declared data collection period.
Notes / Findings
Control Objective
The organization shall document data lineage from source to model input, including all transformation steps and retention policies.
SWT3 Procedure
AI-DATA.3 -- Data Lineage Attestation
Factor Legend
AI-DATA.3: factor_a = lineage stage | factor_b = transformation type | factor_c = retention period (days) (SDK defaults -- confirm with provider)
Test Procedure
  • Trace data lineage from source to model input using AI-DATA.3 anchors.
  • Verify all transformation steps are documented with type and stage information.
  • Confirm retention policies (factor_c) are included and meet regulatory requirements.
Expected Evidence
Complete data lineage covering all transformation steps from raw source to model-ready input. Retention policies documented at each stage.
Sample Verification
sovereign.tenova.io/verify -- paste any AI-DATA.3 anchor token
Pass/Fail Criteria
PASS if: Lineage covers all transformations from source to model input. FAIL if: Transformation steps are undocumented, or lineage has gaps between source and model.
Red Flag
Watch for: AI-DATA.3 lineage records stop at an intermediate data warehouse. Source-to-model traceability has a gap.
Notes / Findings
Control Objective
The organization shall conduct data protection impact assessments and broader societal impact assessments, including environmental and equity dimensions, before deploying high-risk AI systems.
SWT3 Procedures
AI-DPIA.1 -- Data Protection Impact Assessment | AI-IMPACT.1 -- Societal Impact Assessment
Factor Legend
AI-DPIA.1: factor_a = assessment scope | factor_b = risk rating (1=low, 2=medium, 3=high) | factor_c = supervisory authority (identifier or 0 if not applicable)
AI-IMPACT.1: factor_a = assessment scope | factor_b = impact rating (1=low, 2=medium, 3=high) | factor_c = affected population size (SDK defaults -- confirm with provider)
Test Procedure
  • Verify AI-DPIA.1 anchors predate deployment for all systems processing personal data.
  • Check AI-IMPACT.1 covers environmental and equity dimensions in addition to technical risks.
  • Confirm assessments are renewed at the organization's stated reassessment interval.
Expected Evidence
Impact assessments predating deployment for high-risk systems. DPIA covering data protection requirements. Societal impact assessment covering environmental and equity dimensions.
Sample Verification
sovereign.tenova.io/verify -- paste any AI-DPIA.1 or AI-IMPACT.1 anchor token
Pass/Fail Criteria
PASS if: Assessments predate deployment for high-risk systems and cover data protection, environmental, and equity dimensions. FAIL if: No assessment exists for systems affecting individuals, or assessments postdate deployment without justification.
Red Flag
Watch for: AI-DPIA.1 anchor timestamps postdate the first AI-INF.1 inference anchor. The impact assessment was conducted retroactively, not before deployment.
Notes / Findings

Measure Measure Function

The Measure function addresses metrics, testing, and monitoring of AI systems. Entries WP-17 through WP-19 cover performance and fairness metrics, drift and robustness monitoring, and red team and supply chain testing.

Control Objective
The organization shall track quantitative metrics for AI system performance, fairness, and explainability, with defined thresholds and regular cadence.
SWT3 Procedures
AI-FAIR.1 -- Fairness Attestation | AI-PERF.1 -- Performance Monitoring | AI-INF.2 -- Inference Quality | AI-EXPL.1 -- Explainability Attestation
Factor Legend
AI-FAIR.1: factor_a = protected attribute tested | factor_b = disparity ratio (1.0 = no disparity) | factor_c = acceptable threshold
AI-PERF.1: factor_a = performance metric | factor_b = measured value | factor_c = target value
AI-INF.2: factor_a = quality metric | factor_b = measured score | factor_c = minimum acceptable score
AI-EXPL.1: factor_a = explanation method (e.g., SHAP, LIME, attention) | factor_b = confidence score | factor_c = factors cited count (SDK defaults -- confirm with provider)
Test Procedure
  • Request time series of AI-FAIR.1 anchors and verify disparity ratios (factor_b) remain within thresholds (factor_c).
  • Cross-reference AI-PERF.1 measured values (factor_b) with model card performance claims (factor_c).
  • Check AI-INF.2 inference quality metrics for production consistency.
  • Verify AI-EXPL.1 confirms explanations are generated with appropriate methods for the risk level.
Expected Evidence
Fairness metrics tracked over time with thresholds defined. Performance metrics measured against documented targets. Explainability methods documented and operational.
Sample Verification
sovereign.tenova.io/verify -- paste any AI-FAIR.1 or AI-PERF.1 anchor token
Pass/Fail Criteria
PASS if: Metrics are tracked with defined thresholds and regular cadence. FAIL if: Fairness is measured without defined thresholds, or performance metrics are untracked for production systems.
Red Flag
Watch for: AI-FAIR.1 disparity ratios are all exactly 1.000 (no disparity). Perfect fairness is statistically improbable and suggests the metric is not being computed on real data.
Notes / Findings
Control Objective
The organization shall establish baselines, monitor model drift, verify model integrity, and conduct robustness testing for production AI systems.
SWT3 Procedures
AI-BASE.1 -- Baseline Establishment | AI-DRIFT.1 -- Model Drift Detection | AI-MDL.3 -- Model Integrity Verification | AI-ROBUST.1 -- Robustness Testing
Factor Legend
AI-BASE.1: factor_a = baseline metric | factor_b = baseline value | factor_c = establishment date hash
AI-DRIFT.1: factor_a = drift metric | factor_b = observed magnitude | factor_c = configured threshold
AI-MDL.3: factor_a = model identifier | factor_b = integrity hash | factor_c = verification result (1=intact)
AI-ROBUST.1: factor_a = test type | factor_b = perturbation magnitude | factor_c = stability score (SDK defaults -- confirm with provider)
Test Procedure
  • Verify AI-BASE.1 baseline was established before deployment. See Query 4: Temporal Continuity.
  • Check AI-DRIFT.1 monitoring frequency (weekly minimum for production systems).
  • Verify threshold exceedances (factor_b exceeding factor_c) triggered corrective action.
  • Confirm AI-ROBUST.1 adversarial robustness testing is documented.
  • Verify AI-MDL.3 integrity checks confirm model has not been tampered with.
Expected Evidence
Baseline established before deployment. Regular drift monitoring with defined thresholds. Robustness testing documented. Model integrity verified on a regular cadence.
Sample Verification
sovereign.tenova.io/verify -- paste any AI-DRIFT.1 or AI-BASE.1 anchor token
Pass/Fail Criteria
PASS if: Baseline exists, drift is monitored regularly, threshold exceedances trigger corrective response, and robustness is tested. FAIL if: No baseline exists, or drift is unmonitored for production models.
Red Flag
Watch for: AI-BASE.1 baseline exists but no AI-DRIFT.1 anchors reference it. Drift is being measured against an unknown baseline.
Notes / Findings
Control Objective
The organization shall conduct red team assessments and supply chain testing before deployment or on an annual basis, and document findings with severity classification.
SWT3 Procedures
AI-REDTEAM.1 -- Red Team Assessment | AI-SUPPLY.1 -- Supply Chain Assessment
Factor Legend
AI-REDTEAM.1: factor_a = test scope | factor_b = findings count | factor_c = severity distribution (encoded)
AI-SUPPLY.1: factor_a = supplier count assessed | factor_b = risk findings | factor_c = assessment date hash (SDK defaults -- confirm with provider)
Test Procedure
  • Verify AI-REDTEAM.1 anchors predate deployment or occur at least annually.
  • Cross-reference AI-SUPPLY.1 supply chain assessment with AI-SBOM.1 from WP-11.
  • Confirm critical findings from red team are resolved before go-live.
Expected Evidence
Red team assessment with findings classification predating deployment. Supply chain assessment covering all third-party AI components.
Sample Verification
sovereign.tenova.io/verify -- paste any AI-REDTEAM.1 or AI-SUPPLY.1 anchor token
Pass/Fail Criteria
PASS if: Adversarial testing is documented before deployment and supply chain is assessed. FAIL if: No V&V evidence before deployment, or supply chain is unassessed.
Red Flag
Watch for: AI-REDTEAM.1 shows zero findings (factor_b = 0). A red team that finds nothing either had an insufficient scope or the results are being filtered.
Notes / Findings

Manage Manage Function

The Manage function addresses operations, incident response, and post-market monitoring. Entries WP-20 through WP-25 cover model lifecycle, cybersecurity, safety and security controls, access control and revocation, incident response and autonomous operations, and post-market monitoring.

Control Objective
The organization shall register all production AI models with documented approval chains and lifecycle tracking from registration through retirement.
SWT3 Procedure
AI-MDL.1 -- Model Registration
Factor Legend
AI-MDL.1: factor_a = model identifier | factor_b = weight file hash | factor_c = framework/runtime version (SDK defaults -- confirm with provider)
Test Procedure
  • Verify AI-MDL.1 registration anchors exist for every model in production.
  • Check that registration predates the first inference anchor for each model.
  • Confirm the approval chain is documented (cross-reference with AI-GOV.7 governance decisions from WP-08).
Expected Evidence
Registration anchors for all production models. Registration timestamps predating first inference. Approval chain documented through governance decision anchors.
Sample Verification
sovereign.tenova.io/verify -- paste any AI-MDL.1 anchor token
Pass/Fail Criteria
PASS if: All production models are registered with documented approval authority. FAIL if: Unregistered models are operating in production.
Red Flag
Watch for: AI-MDL.1 model registration exists for the primary model but not for the auxiliary models listed in AI-SBOM.1.
Notes / Findings
Control Objective
The organization shall assess AI-specific cybersecurity threats, including prompt injection, model inversion, training data extraction, and adversarial attacks.
SWT3 Procedure
AI-CYBER.1 -- AI Cybersecurity Assessment
Factor Legend
AI-CYBER.1: factor_a = threat category | factor_b = threats assessed count | factor_c = mitigations in place count (SDK defaults -- confirm with provider)
Test Procedure
  • Verify AI-CYBER.1 anchors document AI-specific threats (prompt injection, model inversion, training data extraction).
  • Check assessment currency -- should be less than 6 months old for high-risk systems.
  • Confirm threat categories (factor_a) cover AI-specific vectors, not just general IT threats.
Expected Evidence
Cybersecurity assessment covering AI-specific threat vectors. Current assessment (less than 6 months old). Mitigations documented for identified threats.
Sample Verification
sovereign.tenova.io/verify -- paste any AI-CYBER.1 anchor token
Pass/Fail Criteria
PASS if: AI-specific threat assessment is current (less than 6 months) and covers AI-unique vectors. FAIL if: Assessment covers only general IT threats, or assessment is older than 6 months for high-risk systems.
Red Flag
Watch for: AI-CYBER.1 threat assessment covers only standard IT threats (phishing, malware). AI-specific threats (prompt injection, model inversion, training data extraction) are absent.
Notes / Findings
Control Objective
The organization shall define concrete safety boundaries with active mitigations and maintain security attestation aligned with vulnerability management cadence.
SWT3 Procedures
AI-SAFE.1 -- Safe State Verification | AI-SEC.1 -- Security Attestation | AI-SEC.2 -- Security Posture
Factor Legend
AI-SAFE.1: factor_a = risk scenario | factor_b = mitigation active (1=yes) | factor_c = safe state confirmed (1=yes)
AI-SEC.1: factor_a = security domain | factor_b = attestation status (1=attested) | factor_c = last assessment date hash
AI-SEC.2: factor_a = posture score | factor_b = vulnerability count | factor_c = remediation rate (SDK defaults -- confirm with provider)
Test Procedure
  • Confirm AI-SAFE.1 defines concrete safety boundaries (not vague descriptions) with mitigations active (factor_b = 1).
  • Verify AI-SEC.1 security attestation cadence aligns with the organization's vulnerability management schedule.
  • Check AI-SEC.2 posture score trend over time.
Expected Evidence
Specific safety boundaries with active mitigations and confirmed safe state. Security attestation on a regular cadence. Posture score tracked over time.
Sample Verification
sovereign.tenova.io/verify -- paste any AI-SAFE.1 or AI-SEC.1 anchor token
Pass/Fail Criteria
PASS if: Safety boundaries are defined with active mitigations and security attestation is current. FAIL if: Boundaries are vague or mitigations are inactive, or security attestation has lapsed.
Red Flag
Watch for: AI-SAFE.1 safe state definitions use qualitative language ('best effort') rather than quantifiable boundaries.
Notes / Findings
Control Objective
The organization shall enforce access control for AI systems and maintain a tested revocation procedure for retiring models, revoking outputs, or responding to incidents.
SWT3 Procedures
AI-ACC.1 -- Access Control Attestation | AI-REV.1 -- Anchor Revocation
Factor Legend
AI-ACC.1: factor_a = resource identifier | factor_b = access decision (1=granted, 0=denied) | factor_c = requester identity hash
AI-REV.1: factor_a = revoked anchor fingerprint | factor_b = reason code (0-6) | factor_c = authority identifier
Reason codes: 0=unspecified, 1=model_recall, 2=policy_violation, 3=data_contamination, 4=consent_withdrawal, 5=regulatory_order, 6=error_correction (SDK defaults -- confirm with provider)
Test Procedure
  • Verify AI-ACC.1 confirms active access control for AI system resources.
  • Check AI-REV.1 revocation log for the assessment period. See Query 5: Revocation Audit.
  • Verify revoked anchors are flagged in the public verifier at sovereign.tenova.io/verify.
  • Confirm the revocation procedure has been tested (at least one test revocation or documented tabletop).
Expected Evidence
Active access control with documented decisions. Revocation procedure tested. Revoked anchors flagged in public verifier. Specific reason codes used (not unspecified).
Sample Verification
sovereign.tenova.io/verify -- paste any AI-ACC.1 or AI-REV.1 anchor token
Pass/Fail Criteria
PASS if: Access control is enforced and revocation procedure is tested. FAIL if: No revocation capability exists, or access control is unenforced for AI resources.
Red Flag
Watch for: AI-REV.1 revocations all use reason code 0 (unspecified). Specific reason codes (1-6) are required for meaningful audit trail.
Notes / Findings
Control Objective
The organization shall maintain an AI-specific incident response plan, conduct tabletop exercises, and document autonomous AI system decisions with appropriate oversight.
SWT3 Procedures
AI-IR.1 -- Incident Response Plan | AI-AUTO.1 -- Autonomous Decision Recording | AI-INCIDENT.1 -- Incident Detection
Factor Legend
AI-IR.1: factor_a = plan version | factor_b = last exercise date hash | factor_c = plan scope
AI-AUTO.1: factor_a = decision type | factor_b = confidence score | factor_c = human review required (1=yes)
AI-INCIDENT.1: factor_a = incident category | factor_b = severity (1=low, 2=medium, 3=high, 4=critical) | factor_c = response time (minutes) (SDK defaults -- confirm with provider)
Test Procedure
  • Verify AI-IR.1 confirms an AI-specific incident response plan exists (not just a generic IT IR plan).
  • Check for tabletop exercise evidence (factor_b should indicate a recent exercise date).
  • For autonomous systems, verify AI-AUTO.1 records all autonomous decisions with confidence scores.
  • Review AI-INCIDENT.1 anchors for response time compliance with stated SLA.
Expected Evidence
AI-specific IR plan with documented tabletop exercise. Incident records with severity classification and response times. Autonomous decision records for systems operating without direct human oversight.
Sample Verification
sovereign.tenova.io/verify -- paste any AI-INCIDENT.1 or AI-AUTO.1 anchor token
Pass/Fail Criteria
PASS if: AI-specific IR plan is tested and incidents are documented with severity and response times. FAIL if: No AI-specific IR plan exists, or procedures are untested.
Red Flag
Watch for: AI-INCIDENT.1 response times (factor_c) are consistently under 60 seconds for high-severity incidents. This suggests automated acknowledgment, not human triage.
Notes / Findings
Control Objective
The organization shall conduct post-market monitoring of deployed AI systems, track policy violations, and ensure human review volume is appropriate for the system's risk profile.
SWT3 Procedures
AI-HITL.2 -- Human Review Volume | AI-PMM.1 -- Post-Market Monitoring | AI-VIO.1 -- Policy Violation Detection
Factor Legend
AI-HITL.2: factor_a = review sample size | factor_b = reviews completed | factor_c = rejection rate
AI-PMM.1: factor_a = monitoring metric | factor_b = measured value | factor_c = monitoring cadence (hours)
AI-VIO.1: factor_a = violation type | factor_b = severity (1-4) | factor_c = enforcement action taken (1=yes) (SDK defaults -- confirm with provider)
Test Procedure
  • Verify AI-PMM.1 post-market monitoring is on a regular cadence appropriate for the system's risk tier.
  • Cross-reference AI-VIO.1 violations with AI-GOV.5 enforcement actions from WP-10.
  • Check AI-HITL.2 reviewer volume -- is the sample size appropriate for the decision volume and risk level?
  • Verify violations flow to enforcement (factor_c in AI-VIO.1 = 1 for documented violations).
Expected Evidence
Regular post-market monitoring at documented cadence. Violations detected and flowing to enforcement. Human review volume proportionate to system risk.
Sample Verification
sovereign.tenova.io/verify -- paste any AI-PMM.1 or AI-VIO.1 anchor token
Pass/Fail Criteria
PASS if: Post-market monitoring is active and violations flow to enforcement. FAIL if: No monitoring exists after deployment, or violations are detected but never enforced.
Red Flag
Watch for: AI-PMM.1 post-market monitoring anchors exist monthly but performance metrics (factor values) are identical across reports -- suggesting copy-paste rather than fresh measurement.
Notes / Findings

How to Cite Verified Evidence

When referencing SWT3 anchor evidence in your AI RMF assessment report, use a consistent citation format. The following templates cover individual anchor verification and workpaper-level summaries.

Per-anchor citation

"Per WP-[##], the assessor independently verified SWT3 Witness Anchor [full token] using SHA-256 fingerprint recomputation (FIPS 180-4). The anchor confirms that procedure [ID] was evaluated on [date] with verdict [PASS/FAIL]. Factors: [descriptions with values]. Verification result: fingerprint match confirmed."

Workpaper summary citation

"The assessor reviewed [N] SWT3 witness anchors across [M] procedures covering [date range]. [X] anchors were independently verified via fingerprint recomputation. Coverage gaps: [list or 'none identified']. Workpaper entries WP-01 through WP-25 document findings across all four AI RMF functions (Govern, Map, Measure, Manage). Overall assessment: [assessor's determination]."

Common Assessment Questions

Questions that arise during AI RMF assessments, with guidance on how to proceed.

What if the client uses different factor mappings than this workpaper describes?
The factor legends in this workpaper reflect SWT3 SDK defaults. If the client has customized their factor usage, request their factor mapping document. Verify anchors using the client's stated mappings, and note any deviations from SDK conventions in your findings.

What if factor values are redacted or minimal?
SWT3 supports multiple evidence detail levels (clearing levels 0-3). If factors show minimal information (e.g., all values are 0 or 1), ask the client which clearing level they have configured and whether the restriction is appropriate for the system's risk classification. Note the clearing level in your findings.

What if the organization has very few anchors (new deployment)?
For recently deployed systems, the anchor stream will be short. Evaluate based on the time elapsed since deployment. A system deployed 30 days ago with 30 days of anchors demonstrates continuous monitoring from day one. Adjust your temporal continuity expectations to the deployment date, not a fixed 90-day window.

What if an anchor fails fingerprint verification?
A verification failure means the anchor has been modified since minting, or the verification inputs are incorrect. First, confirm you have the correct inputs (tenant ID, procedure ID, factors, millisecond timestamp). If inputs are correct and verification still fails, this is a significant finding -- document it and request the client's explanation.

How do lifecycle chains (v6.0) relate to AI RMF categories?
Lifecycle chains link multiple anchors into a sequence using a shared cycle_id. For the Govern function, chains document governance decision sequences. For Manage, chains link incident detection through response and resolution. When reviewing lifecycle chains, verify the chain is complete (no missing sequence positions) and that the cycle_id is consistent across all anchors in the chain.

What if the organization is assessed against multiple frameworks simultaneously?
SWT3 procedures are framework-agnostic -- the same AI-FAIR.1 anchor satisfies fairness requirements in NIST AI RMF, ISO 42001, EU AI Act, and SR 11-7. When an organization is assessed against multiple frameworks, you can cite the same anchor evidence in each framework's workpaper with the appropriate framework-specific citation. The crosswalk guides at sovereign.tenova.io/guides map procedures to each framework's specific requirements.

Summary Sheet

Record the assessment result for each workpaper entry. Use this as the summary page for audit file documentation.

EntryCategoryResultEvidence ReferenceInitials
GOVERN FUNCTION
WP-01GOVERN 1.1 Dual-Use and Oversight
WP-02GOVERN 1.2 Technical Environment
WP-03GOVERN 1.3 Supply Chain
WP-04GOVERN 1.4 Audit and Logging
WP-05GOVERN 1.5 Trust and Guardrails
WP-06GOVERN 1.7 Transparency
WP-07GOVERN 2.1 Roles
WP-08GOVERN 2.2 Governance Mechanisms
WP-09GOVERN 4.1 Human Oversight
WP-10GOVERN 6.1 Policy Enforcement
MAP FUNCTION
WP-11MAP 1.1 System Identification
WP-12MAP 2.1 Risk Identification
WP-13MAP 2.3 Fairness and Explainability
WP-14MAP 3.5 Data Governance
WP-15MAP 4.1 Data Lineage
WP-16MAP 5.2 Impact Assessment
MEASURE FUNCTION
WP-17MEASURE 2.5 Performance and Fairness
WP-18MEASURE 2.6 Drift and Robustness
WP-19MEASURE 3.1 Red Team and Supply Chain
MANAGE FUNCTION
WP-20MANAGE 1.3 Model Lifecycle
WP-21MANAGE 2.2 Cybersecurity
WP-22MANAGE 2.3 Safety and Security
WP-23MANAGE 2.4 Access and Revocation
WP-24MANAGE 3.1/3.2 Incident Response
WP-25MANAGE 4.1 Post-Market Monitoring
Lead Assessor Signature / Date
Technical Expert Signature / Date