• qa_deck_100#slide 32 · Socio-technical governance and accountability lineage

RAIDT treats GenAI use as a socio-technical system because organisational risk does not arise from the model in isolation. Across the papers, materially different outcomes are produced by the configured use of a system in context: prompts, retrieved evidence, tool calls, adapters, safety settings, workflow rules, and human review steps all shape what a run produces and how that output is relied upon. For that reason, RAIDT treats the run as the unit of governance. The central governance question is not simply whether a model is documented or capable in the abstract, but whether one situated use can later be reconstructed, reviewed, compared, and challenged.

This framing is socio-technical in the Information Systems sense. Governable use depends on the interaction of technical artefacts with roles, routines, decision rights, and accountability structures. A technically instrumented system can still be weakly governed if no one knows who reviewed the output, what escalation path applied, or whether evidence was retained; equally, strong governance intentions fail if the technical system does not preserve usable traces. RAIDT therefore uses the run-level evidence pack and the score profile across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) to make governance readiness inspectable where risk materialises. On this logic, influence methods as governance interventions matter because they change both behaviour and evidencing. RAIDT governs configured organisational use, not the model alone.

In a public-service eligibility workflow, a caseworker uses a GenAI assistant to interpret an entitlement rule and draft advice for a claimant. The output may sound authoritative, but the real socio-technical system also includes the caseworker's prompt, the approved rule repository, the retrieval layer, the model version, and the supervisor's review decision. If only the answer text is kept, the organisation cannot show which clause was used or whether the advice was checked.

With RAIDT, the run-level evidence pack stores the prompt, model and version identifiers, retrieval snapshot with document hashes, output, and the reviewer note recording whether the advice was accepted or escalated. The score profile can then show strong Auditability and Traceability, or expose weakness if the clause snapshot or human sign-off is missing. The example shows why RAIDT treats GenAI use as socio-technical and why the run becomes the object of governance.

• 17-RAIDT-Sociotechnical_M_v6
• 12-RAIDT_DSR_Theory_M_v8