5D artifact map

What evidence does 5D produce?

Five technical outputs across the action-governance lifecycle. Each one has a defined producer, audience, purpose, and limit.

Read the ownership first

Technical evidence is not a compliance deliverable.

No framework below mandates a 5D artifact by name. These outputs may contribute technical evidence relevant to the cited provision when it applies. Applicability, evidence sufficiency, custody, retention, legal conclusions, and compliance remain with the organisation.

Artifact register

Five outputs. Each with a boundary.

5D creates the records listed here. It does not create the organisation's approvals, impact assessments, reviewer authority, execution outcomes, retention policy, or decommissioning decision.

01 · Test output

Deterministic test report

Produced by: 5D offline harness
Produced when: When declared scenarios are run against the selected release and policy conditions.
Contains: A structured summary of expected and observed results, with optional per-case results and explicit test limitations.

Why teams need it

It makes the tested behavior reviewable before policy is promoted or changed.

Primary users

Agent engineers Security testing Governance reviewers Platform owners

What it does not establish

A pass means the declared expectation matched the observed result.
It does not prove system robustness, security, integration coverage, or test sufficiency.

Relevant framework references

EU AI Act · Test

Articles 9(7) and 15(1), 15(4)

Testing, robustness and cybersecurity

Where this output may help: Deterministic policy tests and explicit failure states can contribute action-level testing evidence.

Limit: Only relevant where these provider duties apply; 5D scenario tests do not establish system accuracy, robustness, cybersecurity or Article 9 test sufficiency.

Official source ↗

NIST AI RMF · Test

MEASURE 1.1–1.2; 2.1, 2.3 and 2.6–2.7

Measurement approaches, test sets, performance and risk evaluation

Where this output may help: Release-specific deterministic tests, controlled scenarios and replay can measure a declared policy under stated conditions.

Limit: This is evidence for selected action-policy behavior, not verified satisfaction of the NIST subcategories or evaluation of the full AI system.

Official source ↗

OWASP Agentic · Test

ASI01 Agent Goal Hijack — mitigation 8; ASI02 Tool Misuse and Exploitation — mitigations 1–4

Red-team scenarios and bounded tool-use controls

Where this output may help: Hostile scenarios can test one technical measure: pre-action policy on covered tool calls.

Limit: 5D does not mitigate either complete ASI risk; effectiveness depends on realistic scenarios, complete interception and host enforcement.

Official source ↗

02 · Policy provenance

Policy-content digest

Produced by: 5D policy engine
Produced when: When policy content is hashed and the digest is linked to a decision record.
Contains: A stable digest of the policy fields that can change a verdict.

Why teams need it

It distinguishes the policy content actually used, even when a human-readable version label is reused.

Primary users

Release engineers Security reviewers Policy owners Platform owners

What it does not establish

The digest is not a signature.
It does not identify or authenticate the approver, prove release authorisation, or establish policy sufficiency.

Relevant framework references

EU AI Act · Approve and version

Articles 11(1) and 17(1)(a)–(d)

Technical documentation and quality-management controls

Where this output may help: Versioned policy, configuration and test artifacts may support one documented change-control record.

Limit: Only relevant to applicable provider duties; 5D neither creates the complete technical documentation nor establishes a quality-management system.

Official source ↗

NIST AI RMF · Approve and version

GOVERN 1.4–1.5 and 2.1–2.3

Risk culture, ongoing review, roles and executive responsibility

Where this output may help: Policy versions and evidence hashes can make one technical approval decision reconstructable.

Limit: 5D does not assign organisational roles, establish risk culture or provide executive accountability.

Official source ↗

ISO/IEC 42001 · Approve and version

Clauses 7.5 and 8.1 — Documented information; operational planning and control

Documented-information and operational-control objectives

Where this output may help: Versionable policy artifacts may support controlled documented information for one runtime control.

Limit: Objective-level titles only; licensed normative text is not reproduced, and 5D does not establish management-system control or conformity.

Official source ↗

03 · Runtime record

Action-level decision record

Produced by: 5D decision logger
Produced when: When logging is configured and a proposed action is evaluated.
Contains: Selected action inputs, policy provenance, rule attribution, disposition, rationale, acting-principal context, and separately recorded outcome history.

Why teams need it

It lets teams reconstruct which policy path evaluated one proposed action and what 5D decided.

Primary users

Agent engineers Security operations Governance and audit Platform owners

What it does not establish

The record is persistent only when logging is configured and available.
The host supplies identity assertions, review evidence, execution attempts, and outcomes. Completeness, custody, and retention remain host responsibilities.

Relevant framework references

EU AI Act · Observe and review

Articles 12(1)–(2), 14 and 26(2), 26(5)

Record-keeping, human oversight and deployer monitoring

Where this output may help: Action receipts and a human-review route can contribute selected technical evidence for a covered path.

Limit: Article 26 applies to deployers of applicable high-risk AI systems. 5D does not determine role or applicability, nor establish required log scope, completeness, custody, retention or oversight sufficiency.

Official source ↗

Singapore Agentic AI · Observe and review

Dimensions 2 and 3 — Enable meaningful human accountability; implement technical controls and processes

Human checkpoints, logging and monitoring

Where this output may help: A policy-selected hold and action receipt can support one human checkpoint and record the resulting disposition.

Limit: The organisation owns reviewer identity, authority, competence, escalation, monitoring coverage and accountability.

Official source ↗

NIST AI RMF · Observe and review

MAP 3.5; MEASURE 2.4 and 3.1; MANAGE 4.1

Human oversight, measurement documentation and post-deployment monitoring

Where this output may help: Separate policy, review and outcome states can contribute action-level evidence to oversight and monitoring.

Limit: 5D does not establish organisation-wide monitoring, assessor independence, reviewer sufficiency or incident response.

Official source ↗

04 · Operational event

Optional operational event stream

Produced by: 5D NDJSON event channel
Produced when: When the optional event path is configured.
Contains: Risk-decision, budget-intervention, and identity-required-denial events for downstream correlation.

Why teams need it

It exposes selected runtime events to the organisation's monitoring and operational workflows.

Primary users

Security operations Platform operations

What it does not establish

The stream is optional and best-effort.
An event-write failure does not stop the action path, and this stream is not a complete audit trail or incident record.

Relevant framework references

NIST AI RMF · Observe and review

MAP 3.5; MEASURE 2.4 and 3.1; MANAGE 4.1

Human oversight, measurement documentation and post-deployment monitoring

Where this output may help: Separate policy, review and outcome states can contribute action-level evidence to oversight and monitoring.

Limit: 5D does not establish organisation-wide monitoring, assessor independence, reviewer sufficiency or incident response.

Official source ↗

Singapore Agentic AI · Observe and review

Dimensions 2 and 3 — Enable meaningful human accountability; implement technical controls and processes

Human checkpoints, logging and monitoring

Where this output may help: A policy-selected hold and action receipt can support one human checkpoint and record the resulting disposition.

Limit: The organisation owns reviewer identity, authority, competence, escalation, monitoring coverage and accountability.

Official source ↗

05 · Integrity result

Decision-chain verification result

Produced by: 5D decision-log verifier
Produced when: When an operator runs the chain check against a decision log.
Contains: A result stating whether chained decision-time fields recompute and, if not, the first failed row.

Why teams need it

It can reveal a break in the local chain over the decision-time fields that 5D covers.

Primary users

Security assurance Governance and audit

What it does not establish

Chain verification is evidence, not proof of immutability or authoritative custody.
Mutable outcome history is outside the chain, and a writer with local access can rewrite the tail.

Relevant framework references

EU AI Act · Observe and review

Articles 12(1)–(2), 14 and 26(2), 26(5)

Record-keeping, human oversight and deployer monitoring

Where this output may help: Action receipts and a human-review route can contribute selected technical evidence for a covered path.

Official source ↗

NIST AI RMF · Observe and review

MAP 3.5; MEASURE 2.4 and 3.1; MANAGE 4.1

Human oversight, measurement documentation and post-deployment monitoring

Where this output may help: Separate policy, review and outcome states can contribute action-level evidence to oversight and monitoring.

Limit: 5D does not establish organisation-wide monitoring, assessor independence, reviewer sufficiency or incident response.

Official source ↗

ISO/IEC 42001 · Observe and review

Clause 9 — Performance evaluation

AI management-system performance-evaluation objective

Where this output may help: Action receipts may be one technical input to an organisation's monitoring and review process.

Limit: Objective-level reference only; licensed normative text is not reproduced, and receipts do not establish audit, management review or conformity.

Official source ↗

Lifecycle ownership map

Who supplies what at every stage.

A blank 5D output is intentional. Some lifecycle duties belong to the host or organisation and should not be relabelled as product evidence.

01
Design layer
Define
What may this agent do?

5D produces
No standalone 5D artifact claimed at this stage.

Host supplies
The action schema, tool and destination vocabulary, proposed context, and policy inputs.

Organisation owns
Use-case classification, impact assessment, acceptable autonomy, policy intent, and legal applicability.
02
Evidence layer
Test
Does the policy behave as intended?
5D produces
- Test report
Host supplies
Representative scenarios, realistic inputs, negative controls, and declared expectations.

Organisation owns
Test sufficiency, acceptance criteria, independent assurance, remediation, and promotion decisions.
03
Change-control layer
Approve and version
Which policy and evidence are approved for use?
5D produces
- Policy digest
Host supplies
The policy and configuration whose content is hashed.

Organisation owns
Policy ownership, approval record, reviewer authority, release authorisation, and segregation of duties.
04
Deployment contract
Bind and deploy
Which approved control covers each execution path?

5D produces
No standalone 5D artifact claimed at this stage.

Host supplies
Gateway configuration, authenticated context, and the covered action entry points.

Organisation owns
Integration coverage evidence, secure deployment, credential controls, alternate-path prevention, rollback, and recovery.
05
Runtime core
Enforce at action time
May this proposed action cross the boundary?
5D produces
Host supplies
The proposed action, identity and authority context, destination, purpose, and policy binding.

Organisation owns
Execution authority, complete interception, context accuracy, downstream enforcement, and prohibited-path prevention.
06
Evidence and oversight layer
Observe and review
What happened, and can the decision be reconstructed and challenged?
5D produces
Host supplies
Human-review evidence, reviewer identity, execution attempt, and execution outcome.

Organisation owns
Authoritative custody, retention, monitoring, reviewer competence, incident response, notices, redress, and audit.
07
Governed-change layer
Change or retire
Should policy, integration, or agent remain in service?
5D produces
- Test report
Host supplies
Changed policy, new scenarios, observed incidents, and the intended release boundary.

Organisation owns
Reassessment, reapproval, corrective action, rollback, post-market monitoring, decommissioning, and residual-data handling.