• workshop_table17#tag-band S10–S12 · 165–180 min

The project is supported by three mutually reinforcing forms of evidence across the cited papers. First, the foundations paper establishes the core problem and artefact logic: governance failures materialise in configured uses, so the run as the unit of governance and the run-level evidence pack are needed to make one use event reconstructable, reviewable, and comparable. Second, the academic-logic paper consolidates programme-level support by showing how the artefact becomes measurable and usable in practice: the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability), the score profile, implementation modes, policy translation, and a cross-domain empirical programme using repeated runs, calibration, and multiple influence configurations. Third, the design-theory paper explains why these elements belong together conceptually: RAIDT is a bounded, mechanism-based mid-range design theory in which influence methods as governance interventions shape both behaviour and evidence quality.

The boundaries are equally important. RAIDT is for organisational generative AI uses whose outputs influence decisions, records, communications, or services; it assumes bounded evidence can be captured under suitable access, privacy, and retention controls; and it is risk-proportionate rather than universal. It is not a guarantee of correctness, not a substitute for law or professional judgement, and not a general theory of all AI governance. The supervision roadmap should therefore be organised centre outward: define the material run, specify the run-level evidence pack, score the five pillars using anchors 1=missing / 3=partial / 5=audit-ready, examine interventions, choose implementation depth, translate to policy, and then extend into the empirical programme. That roadmap keeps supervision coherent because every later claim remains anchored to the same evidentiary object.

In a cybersecurity programme, a supervisor might begin with one alert-triage run and ask whether the recommendation can actually be reconstructed. The team would gather the prompt, enabled tools, retrieved threat-intelligence snapshot, output, hashes, and analyst review notes into a run-level evidence pack. The next step would be to score the five pillars and identify whether the main weakness is Dependability, Auditability, or Traceability.

Only after that inner work would the roadmap widen. The team could test whether a stacked configuration improves the score profile, decide whether orchestration-layer logging is needed, map the evidence grammar to assurance or policy expectations, and then design a broader empirical programme across repeated incidents. In that sequence, supervision remains disciplined, cumulative, and evidence-led rather than rhetorical.

• 08-RAIDT_Foundations_M_V50
• 11-RAIDT_Academic_Logic_M_v11
• 12-RAIDT_DSR_Theory_M_v8