• qa_deck_100#slide 91 · Standards and policy interoperability

RAIDT operationalise EU AI Act duties at run level by specifying what evidence must exist for one material use of generative AI. Rather than treating compliance as a policy statement at model or programme level, RAIDT defines the run as the unit of governance. For each important run, the run-level evidence pack captures the run identifier and timestamp, prompt or template version, model and tool configuration, retrieval query and snapshot where relevant, output and output hash, and the human or automated checks applied. This makes AI Act duties inspectable at the point where organisational risk actually materialises.

The operational logic is then interpreted through the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) and a score profile using anchors 1=missing / 3=partial / 5=audit-ready. Responsibility records named owners, decision rights, escalation playbooks, override narratives, and reviewer sign-off, which supports risk management and human oversight. Auditability and Traceability require versioned prompts, time-stamped logs, hashes, provenance chains, retention rules, and verifier notes, which support documentation and record-keeping. Interpretability requires audience-appropriate explanation sheets and usability checks, which support transparency duties. Dependability requires repeat-run testing, thresholds, rollback planning, and change-control evidence, which support robustness expectations. The papers also emphasise influence methods as governance interventions: prompting, retrieval augmentation, LoRA or PEFT, and preference-based alignment change behaviour and therefore change what evidence must be captured. RAIDT makes those changes governable rather than implicit.

In a public-service workflow, a case worker uses a generative AI tool to explain eligibility rules for a benefit claimant. Under RAIDT, the organisation does not keep only the final answer. It stores the full prompt, model version, retrieval query, snapshot of the policy text used, document identifiers, hashes, output hash, and reviewer decision in the run-level evidence pack. Because the workflow uses retrieval augmentation, the source snapshot is preserved as part of the governance record rather than treated as an optional technical detail.

The score profile can then show whether the run was audit-ready. If the retrieval snapshot is missing, Auditability and Traceability fall. If the explanation is too vague for the claimant, Interpretability falls. If there is no named reviewer or escalation path, Responsibility falls. That is how RAIDT turns AI Act duties into concrete run-level controls rather than abstract organisational promises.

• 10-RAIDT_Policy_Pathways_M_V50
• 16-RAIDT-Audit-Accountability_M_v05