• workshop_dense_100#slide 76

In RAIDT, semi-automated implementation means that evidence capture is partly systematised while governance judgement remains human-led. The empirical paper defines the mode in operational terms: the system logs metadata and the model drafts a structured evidence summary, but a human verifies the result. This matters because RAIDT is built around run as the unit of governance. What must be governed is not a general model description, but one configured use in context, reconstructed through a run-level evidence pack and assessed through a score profile across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). Semi-automation therefore sits between manual pilots and fuller orchestration by reducing clerical effort without weakening reviewability.

Its importance is practical and theoretical. Practically, the papers show that many evidence tasks are objective enough to support automation, especially completeness checks for Auditability and Traceability. At the same time, the literature reviewed in RAIDT shows why raw traces, documentation artefacts, or fluent model summaries are not sufficient on their own; they must be assembled into a bounded governance object that organisational reviewers can inspect. Semi-automation helps do that consistently, while still reserving responsibility-sensitive judgements for human reviewers. It also lowers SME demand because, after calibration, trained non-specialist reviewers can handle many evidence-presence checks and escalate harder cases. In RAIDT terms, semi-automation matters because it enables evidence-centred governance at operational scale, supports influence methods as governance interventions, and makes the difference between anchors 1=missing / 3=partial / 5=audit-ready visible at the level of each run rather than as a vague programme claim.

Consider cybersecurity incident triage. A security operations team uses GenAI to draft an incident summary from analyst notes, tool outputs, and retrieved internal playbooks. Under semi-automated RAIDT, the system records the run ID, prompt template, model version, enabled tools, retrieved playbook snapshot, output hash, and any automated checks. The model then drafts an evidence summary that points to those logged artefacts and suggests whether auditability and traceability appear complete.

An analyst verifies the summary, checks whether uncertainty is stated where evidence is incomplete, and decides whether the draft is safe to use in escalation. This matters because the team gains a reconstructable run-level evidence pack for a high-pressure workflow without replacing expert judgement about operational risk.

• 09-RAIDT_Empirical_M_V50.docx
• 18-RAIDT-Technical-Foundation_M_v04