Focus Human Review on the Actions That Need It
Resolve explicit rules in policy and hold only the actions your policy routes for an authorised decision.
What this path changes
The host can resolve explicit cases through policy and present only policy-selected actions to an authorised reviewer, while keeping policy and reviewer decisions separate.
The action, boundary, and evidence
Who owns the risk
Governance owners, operations leaders, security teams, and engineers responsible for consequential action-review queues.
When the gate runs
An action cannot be permitted by an explicit rule, or policy requires an authorised person to decide before execution.
How to configure assurance
Use deterministic rules for explicit permit and block cases. Route only policy-selected holds to optional semantic evidence or an authorised reviewer; neither may clear an explicit block.
- Approve a consequential payment or account change above a declared threshold.
- Review an external action whose purpose or destination remains uncertain.
- Hold an action when required evidence is missing, malformed, or unavailable.
- A test matrix covering permit, block, hold, missing evidence, review expiry, and reviewer rejection.
- A record that keeps the policy disposition, reviewer decision, execution attempt, and downstream outcome in separate fields.
Next: Test a policy-selected review path with demonstration actions before connecting a live tool.
Follow a matching proposed action through the gate.
Synthetic tested action under the pinned 5D 0.7.0 illustrative policy.
Send a policy exception and refund promise
Generated guidance contradicts the applicable policy and appears authoritative. Under the pinned illustrative policy, the expected result is to hold the action until the host records the required review.
- Tool
- message.send
- Target
- individual traveler
- Purpose
- resolve request
5D suppliesAn ORANGE hold signal before sending the commitment.
Host ownsPolicy retrieval, content review, reviewer authority, and final send.
Coverage limit5D does not determine the correct refund policy.
Where this control fits in the governance lifecycle.
These references identify possible technical-evidence contributions. They do not mean every framework applies to every deployment.
This mapping identifies possible technical-evidence contributions. It is not legal advice, certification, conformity assessment, readiness scoring, a compliance conclusion, or proof that an AI deployment is safe. Applicability, governance, integration coverage, and deployment outcomes remain the organisation's responsibility.
Enforce at action time
May this proposed action cross the boundary?
Articles 9(5)–(6), 14 and 15(1), 15(4)Risk treatment, human oversight, robustness and cybersecurity
What 5D helps with: Pre-action prohibitions, routing and a policy-selected human hold can implement one technical risk-treatment measure.
What 5D does not establish: Only relevant where these duties apply to the actor and system; 5D does not establish the wider risk system, oversight design, robustness or cybersecurity.
Official source ↗MANAGE 1.2–1.3Prioritise and document risk treatment
What 5D helps with: Permit, hold and block policy can implement one documented treatment for selected action risks.
What 5D does not establish: 5D does not prioritise organisation-wide risk or determine risk acceptance and treatment sufficiency.
Official source ↗Observe and review
What happened, and can the decision be reconstructed and challenged?
Section 2 — Transparency and explainability; accountability and governance; contestability and redressTraceable decisions, accountability and challenge
What 5D helps with: Reconstructable action records can support internal explanation and challenge of one policy decision.
What 5D does not establish: 5D does not create affected-person notices, complaint channels, remedies, complete system explanations or regulator-specific accountability.
Official source ↗Dimensions 2 and 3 — Enable meaningful human accountability; implement technical controls and processesHuman checkpoints, logging and monitoring
What 5D helps with: A policy-selected hold and action receipt can support one human checkpoint and record the resulting disposition.
What 5D does not establish: The organisation owns reviewer identity, authority, competence, escalation, monitoring coverage and accountability.
Official source ↗Inspect the policy routing signal
from fivedrisk import classify_tool_call, score, load_policy
policy = load_policy("policy.yaml")
action = classify_tool_call("Bash", {"command": "deploy"}, policy)
result = score(action, policy)
print(result.routing.approval_required)Written for fivedrisk==0.7.0. Verify against the immutable source and public documentation before use.
Resolve explicit cases first
Encode known permissions, prohibitions, thresholds, destinations, and authority requirements in versioned policy so clear cases do not depend on a reviewer or model.
Let policy select the review queue
Hold only the actions whose context or authority cannot be resolved by the declared rules, or that policy explicitly reserves for an authorised human decision.
Preserve the two decisions
Record the original policy disposition and the later reviewer choice separately, including expiry, rejection, unavailability, and whether execution was attempted.
What this page does not establish
- 5D does not supply a universal review interface or decide who is authorised to approve.
- A smaller review queue does not prove lower reviewer fatigue, fewer errors, or a correct final decision.
- Routing quality depends on complete action context, suitable policy, and coverage of every relevant execution path.