• integrated_82#Q1.5

RAIDT is trying to solve the problem of reconstructable accountability for individual GenAI uses in organisational work. Across the papers, the central diagnosis is that many responsible-AI approaches are principle-heavy but operationally thin. They state what organisations ought to value, yet they do not specify the minimum evidence that must exist for each use so that an output can later be reconstructed, reviewed, contested, and improved. This becomes especially serious for GenAI because outputs are shaped not only by the model but by run-time configuration, retrieved context, tool chains, and human review arrangements.

The practical problem, then, is not merely whether a model is capable or documented. It is whether an organisation can later show what happened in a particular run: what prompt and template were used, which deployment and parameters were active, what sources informed the output, what checks were applied, and what oversight decision followed. RAIDT addresses that problem by treating the run as the unit of governance and by making the run-level evidence pack the core governance object. The project therefore reframes responsible GenAI governance as an evidence and measurement problem. Its contribution is not a claim that a new model solves trust, but that a new governance measurement object makes responsible use inspectable through the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) and a comparable score profile.

An HR team uses GenAI to draft shortlist justifications and interview questions. Months later, an unsuccessful candidate challenges the process. The employer may have a fairness policy and a vendor transparency document, yet still be unable to show which criteria version, prompt template, adapter configuration, and reviewer checks shaped that particular shortlist. The core organisational problem is therefore accountability for the individual use event, not just general assurance about the system.

With RAIDT, the contested shortlist is linked to a run-level evidence pack containing the prompt version, criteria version, model deployment ID, any PEFT or LoRA identifier, the output, and the human review record. That turns an abstract governance promise into a reviewable case file.

• 08-RAIDT_Foundations_M_V50
• 11-RAIDT_Academic_Logic_M_v11