Statements about working behavior link to a source, release, or test.
What supports these claims
See what the evidence supports and what it does not.
Separate shipped behavior from demonstrations and framework mappings. Each record names its source, strength, and limits.
Evidence library
One topic. One clear page.
Every page separates the strength of the evidence from the kind of material shown.
Authority
Who can permit an action?
Policy, observer, and reviewer roles at the action boundary.
OperationsFailure, retry, and unknown outcomes
Phase-specific states without exactly-once claims.
PrivacyField-level data boundary
What crosses to the observer, stays local, and enters evidence.
ReconstructionIllustrative receipt
A versioned schema, example receipts, stable serialization, and ordered events.
IntegrationVerified entry-point boundaries
What release 0.7.0 implements, what remains partial, and what the host must own.
OversightFramework mappings
Evidence concepts, primary sources, and limitations.
Research lineageWhy these five dimensions
The founder origin, research lineage, and source boundaries behind DTREA.
Decision record
See why an action was permitted, reviewed, or blocked.
The demo record shows the proposed action, acting identity, policy reference, optional AI observation, policy decision, optional human review, and simulated execution status.
The website demonstrates an architecture without running an engine, model, or reviewer.
The links show relevant oversight concepts. They do not establish compliance.
Full framework breakdown
Control-by-control evidence boundaries.
Each mapping states what the mechanism contributes and what remains an organisational or legal responsibility.
Regulation (EU) 2024/1689
Primary source ↗Binding EU law whose obligations depend on the system, actor role, use, classification, and applicable date. The map focuses on technical evidence relevant to lifecycle risk management, logging, oversight, and robustness.
Official Journal text, 2024/1689 · verified 2026-10-06 · lawPartial technical contribution
Articles 9 and 15 — Risk management and robustness
Versioned policy, scenario tests, replay, and explicit failure handling can contribute action-level evidence.
A runtime gate is one control inside the provider or deployer's wider risk, quality, and post-market system.Inspect evidence ↗Partial technical contribution
Article 12 — Record-keeping
A decision record can bind the proposed action, declared identity context, policy version, evidence route, disposition, and separately declared outcome states.
Required logging scope, completeness, custody, retention, access, and integrity remain deployment responsibilities.Inspect evidence ↗Partial technical contribution
Article 14 — Human oversight
A policy-selected route can keep an action stopped while the host obtains and records a human disposition.
5D does not establish that the oversight design, reviewer authority, competence, or organisational measures are sufficient.Inspect evidence ↗UK AI regulatory principles
Primary source ↗Five cross-sector principles for regulators to interpret within their remits: safety, security and robustness; transparency and explainability; fairness; accountability and governance; and contestability and redress.
Initial guidance for regulators, 6 February 2024 · verified 2026-10-06 · regulatory-principlesPartial technical contribution
Section 2 — Safety, security and robustness
Pre-action policy and explicit failure states can contribute evidence that a selected action path applies declared controls.
The integrating organisation retains system-level safety, cyber resilience, model evaluation, and sector-specific duties.Inspect evidence ↗Partial technical contribution
Section 2 — Accountability and governance
Versioned decisions can identify policy ownership, the evaluated action, the evidence route, and the declared disposition.
Legal responsibility, governance structure, supply-chain accountability, and assurance remain organisational and regulator-specific.Inspect evidence ↗Contextual anchor
Section 2 — Transparency and explainability; contestability and redress
Reconstructable records can support explanation and internal challenge of an action-level policy decision.
5D does not create affected-person notices, complaint channels, remedies, or a complete explanation of the AI system.Inspect evidence ↗Singapore Model AI Governance Framework for Agentic AI
Primary source ↗Agent-specific guidance on bounding risks and powers, meaningful human accountability, technical controls across the lifecycle, and end-user responsibility.
Updated framework, 20 May 2026 · verified 2026-10-06 · voluntary-frameworkPartial technical contribution
Dimension 1 — Assess and bound the risks
Action envelopes and explicit policy can make tools, destinations, data classes, autonomy context, and red lines inspectable.
Use-case selection, impact assessment, acceptable autonomy, and organisational risk decisions remain with the deployer.Inspect evidence ↗Partial technical contribution
Dimension 2 — Enable meaningful human accountability
Policy can hold selected actions for a host-owned human decision without transferring execution authority to a model.
The host owns checkpoint design, reviewer authentication, competence, action binding, escalation, and accountability.Inspect evidence ↗Partial technical contribution
Dimension 3 — Implement technical controls and processes
5D supplies deterministic policy primitives, tests, and action-entry points that can be composed into a covered path.
Complete path coverage, surrounding security controls, monitoring, and safe operation are not established by the package.Inspect evidence ↗Contextual anchor
Dimension 4 — Enable end-user responsibility
Action-level evidence may help an integrating product communicate selected control outcomes to its users.
5D does not design the user experience, establish user understanding, allocate responsibility, or discharge end-user governance duties.Inspect evidence ↗NIST AI Risk Management Framework 1.0
Primary source ↗A voluntary risk-management structure using Govern, Map, Measure, and Manage. NIST states that AI RMF 1.0 is under revision.
AI RMF 1.0 plus NIST AI 600-1 Generative AI Profile · verified 2026-10-06 · voluntary-frameworkPartial technical contribution
GOVERN 1.3–1.5, 2.1–2.3; MAP 1.1–1.6, 3.3–3.5
Agent scope, authority, tools, destinations, policy ownership, and action context can be made explicit.
Organisation-wide governance, inventory, accountability, risk appetite, and impact mapping sit outside the runtime gate.Inspect evidence ↗Partial technical contribution
MEASURE 1.1–1.2; 2.1, 2.3–2.7; 3.1
Deterministic tests, controlled scenarios, and replay can evaluate a declared policy under stated release conditions.
Results remain specific to the tested integration, corpus, configuration, version, and conditions.Inspect evidence ↗Partial technical contribution
MANAGE 1.2–1.3, 2.4, 3.1 and 4.1–4.3
Runtime routing and decision evidence can support treatment, review, and governed policy change.
5D does not replace risk acceptance, incident management, corrective action, or retirement processes.Inspect evidence ↗ISO/IEC 42001:2023
Primary source ↗An organisation-wide AI management-system standard. This public mapping stays at objective level and does not reproduce or claim conformity with the normative text.
Edition 1, published December 2023 · verified 2026-10-06 · standardPartial technical contribution
Clauses 6 and 8.1 — Planning; operational planning and control
Versioned action policy and controlled promotion can make a runtime control inspectable.
5D does not establish or certify an AI management system, its leadership, planning, support, or complete operational controls.Inspect evidence ↗Contextual anchor
Clauses 9 and 10 — Performance evaluation and improvement
Decision records and reviewed outcomes can feed a governed change process.
Monitoring, internal audit, management review, corrective action, and continual improvement remain organisational processes.Inspect evidence ↗OWASP Top 10 for Agentic Applications 2026
Primary source ↗A threat and control overlay for agentic systems, not law, certification, or a complete security programme.
2026 edition · verified 2026-10-06 · security-guidancePartial technical contribution
ASI02 Tool Misuse and Exploitation — mitigations 1, 2 and 4
Pre-action policy can block, hold, or permit a proposed tool call before the integrated side effect.
Effectiveness depends on complete action-path interception, accurate context, and host enforcement.Inspect evidence ↗Partial technical contribution
ASI03 Identity and Privilege Abuse
Declared identity, authority, destination, and tool privilege can be inputs to deterministic policy.
5D records caller-supplied context; it does not authenticate identity, validate grants, or secure credentials.Inspect evidence ↗Partial technical contribution
ASI02 Tool Misuse and Exploitation — mitigation 8
Decision records can preserve the policy route and separately declared execution outcome.
Record export does not establish completeness, immutability, attribution, monitoring, or incident readiness.Inspect evidence ↗Source manifest
Framework versions and authoritative links.
The status describes this website's mapping, not certification.
- Regulation (EU) 2024/1689Official Journal text, 2024/1689 · supports evidence for · partial · owning control: policy and action-evidence boundary · verified 2026-10-06Authoritative source ↗
- UK AI regulatory principlesInitial guidance for regulators, 6 February 2024 · supports evidence for · partial · owning control: lifecycle accountability record · verified 2026-10-06Authoritative source ↗
- Singapore Model AI Governance Framework for Agentic AIUpdated framework, 20 May 2026 · supports evidence for · partial · owning control: agent lifecycle boundary · verified 2026-10-06Authoritative source ↗
- NIST AI Risk Management Framework 1.0AI RMF 1.0 plus NIST AI 600-1 Generative AI Profile · maps to control objectives · partial · owning control: risk lifecycle record · verified 2026-10-06Authoritative source ↗
- ISO/IEC 42001:2023Edition 1, published December 2023 · maps to control objectives · partial · owning control: AI management-system evidence · verified 2026-10-06Authoritative source ↗
- OWASP Top 10 for Agentic Applications 20262026 edition · maps to control objectives · partial · owning control: agent action boundary · verified 2026-10-06Authoritative source ↗