AI decisions are rarely governed by one policy or one institution. Legal rules, technical standards, payment rails, platform code, model inference, recommendation systems, learned environments and human authority may all shape the final outcome.
NSDM evaluates the complete decision stack: which layer made the decisive choice, whether the system still performs the role that was approved, whether evidence and authority are sufficient, and whether the affected person can understand and contest the result. The first Workbench milestone now implements one narrow part of that stack: authenticated identity and Case-initiation authority are checked separately before tenant-scoped Case state is created.
| State | Meaning | Required behaviour |
|---|---|---|
| governance_clear | Authority, policy, responsibility, jurisdiction, and escalation route are known. | Proceed only if evidence and consequence thresholds also pass. |
| governance_ambiguous | Policy, authority, responsibility, or jurisdiction is unclear. | Ask, abstain, or escalate. |
| policy_allowed | The action is explicitly permitted within scope. | Record scope and evidence; continue with monitoring. |
| policy_forbidden | The action is explicitly prohibited. | Refuse or stop and log the rule. |
| authority_missing | The system or operator lacks decision authority. | Escalate to the accountable role. |
| human_review_required | Policy or consequence requires a human decision. | Prepare an evidence packet and route for review. |
| compliance_sensitive | The action touches regulated, contractual, or rights-sensitive obligations. | Apply logging, retention, and approval controls. |
| high_risk_domain | Error could affect safety, rights, finance, health, trade, or infrastructure. | Use higher evidence thresholds and mandatory escalation. |
| audit_required | The decision must retain an inspectable trace. | Store claim, sources, model/version, policy, decision, and reviewer trail. |
| functional_role_drift | The system is now shaping a different decision or carrying a different risk from the one originally approved. | Reopen review before continued consequential use. |
The 2026 amendment changes timelines, proportionality and several control details. It does not remove the need to prove what a system is doing, which rules apply, or whether safeguards work in practice.
Annex III high-risk rules move to 2 December 2027 and Annex I product-integrated high-risk rules to 2 August 2028. NSDM therefore records the applicable timeline, transition state and whether a legacy deployment is operating under a temporary regime.
Systems intended to generate or manipulate certain non-consensual intimate or child-abuse material are prohibited, including where reasonable and adequate safeguards against foreseeable misuse are absent.
The amended language focuses on supporting staff competence. NSDM requires role-delimited evidence, dated competence records, supervision boundaries and reassessment for volatile domains.
Expanded bias-detection permissions do not create a general licence to process special-category data. The decision record must show strict necessity, purpose limitation, safeguards and accountable approval.
Delayed application for certain legacy obligations does not eliminate the need to know which model produced an output, whether content requires marking, and whether the system changed after approval.
Scaled penalties and simplified modalities should produce risk-proportionate evidence rather than paperwork copied from large-enterprise programmes.
A process can remain intact while the system’s practical role changes. NSDM therefore governs both procedural continuity and functional continuity.
Did a provider update materially alter ranking, generation, refusal, memory, retention or tool-use behaviour?
Has a support tool become the de facto decision because humans routinely defer to it?
Has summarisation become screening, recommendation become exclusion, or simulation become certification?
Did a new data source, agent, vendor, API or downstream consumer alter the risk and decisional context?
Is the same nominal owner still able to understand, override and accept responsibility for the system’s current role?
Material change creates a new evidence state and may reopen approval, even where the original paperwork remains valid.
Practical authority is distributed across several layers. A decision may appear valid inside each individual layer and still be unjustified when the layers conflict, fail to interoperate, or leave responsibility unassigned.
Defines rights, duties, jurisdiction, liability, prohibited conduct, timelines and appeal rights.
Defines formats, identifiers, taxonomies, interoperability and which evidence can be represented or exchanged.
Determines whether a transaction can execute, under which identity and compliance conditions, and whether it is reversible.
Implements the actual defaults, workflows, exclusions, routes, interfaces and user-facing consequences.
Produces classifications, rankings, risk scores, generated claims, recommendations, learned environments and uncertainty estimates.
Assigns approval, override, review, appeal, remedy and liability to identifiable people and organisations.
A legal right may not exist in the technical data model. A platform may implement a narrower rule than the standard requires. A model may produce a valid inference from information unavailable at the prediction moment. A plausible simulation may not represent the physical world faithfully enough for action.
Recommendation is a form of decision-making because ranking determines visibility, exposure, exclusion and behavioural influence.
Record whether the system optimises engagement, revenue, user welfare, diversity, safety or a weighted combination.
Expose which items or people were eligible, excluded or commercially boosted before ranking began.
Measure popularity reinforcement, exposure inequality, filter effects, novelty collapse and long-tail suppression.
Provide practical options to modify personalisation, correct inferred interests or access an alternative ranking mode.
Reduce intensity, abstain or escalate where personalisation exploits addiction, distress, minors or material asymmetries.
Allow affected users or providers to challenge decisive factors, exclusions and materially harmful ranking outcomes.
| State | Meaning | Required behaviour |
|---|---|---|
| jurisdiction_ambiguous | More than one jurisdiction may control the decision, data, actor or transaction. | Pause and resolve applicable authority. |
| cross_layer_conflict | Two or more governance layers produce incompatible requirements or permissions. | Abstain or escalate; do not silently choose a layer. |
| interoperability_failure | Systems cannot reliably exchange or interpret the required data, evidence or control. | Block automation until the boundary is repaired. |
| semantic_mismatch | The same field, category or rule has different meanings across layers. | Request clarification and preserve the mapping. |
| code_policy_divergence | Runtime code does not implement the declared rule or obligation. | Pause deployment and generate a critical defect record. |
| accountable_owner_missing | No identifiable person or organisation owns the decision and its consequences. | Do not permit consequential action. |
| contestability_missing | The affected party has no practical way to challenge the decision. | Require a contestability route before deployment. |
| independent_review_missing | Review exists but is controlled by the same system or team that made the decision. | Route to an independent reviewer. |
| remedy_unavailable | An error can be identified but not corrected, reversed or compensated. | Raise the evidence threshold or prohibit the action. |
| liability_unassigned | Responsibility for harm or loss is not contractually or institutionally allocated. | Escalate before execution. |
| irreversible_action_high_risk | The action cannot be readily reversed and carries material consequences. | Require stronger evidence and mandatory human approval. |
| simulation_domain_mismatch | A learned environment is being used outside the conditions for which its validity was established. | Block action or require field validation. |
No consequential automated decision should proceed without a minimum, standardised route for transparency, challenge and remedy.
The affected person should be told that an automated or algorithmically mediated decision materially influenced the outcome.
The responsible system, organisation, evidence and major rules should be identifiable in meaningful terms.
The affected party needs a practical route to dispute errors, missing context, inappropriate authority or unjustified exclusion.
Review should not be limited to the same model, workflow or team that produced the original decision.
A review process without the power to change an unjust outcome is not sufficient accountability.
Evidence, versions, authority, actions, reviewers and outcomes must remain inspectable.