• workshop_dense_100#slide 78

Gating in RAIDT means that a run or workflow does not move forward unless the recorded evidence and scores meet the minimum standard set for that context. It is therefore a governance control built on the run-level evidence pack and the score profile, not a simple approval of a model family. The relevant judgement concerns the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability), assessed using anchors 1=missing / 3=partial / 5=audit-ready. Because RAIDT treats the run as the unit of governance, the gate evaluates one configured use in its actual organisational setting, including prompts, configuration, retrieved context, outputs, checks, and oversight.

A simple example is an HR shortlist-justification workflow. If the organisation cannot show which prompt template, criteria version, model deployment, or adapter configuration produced the recommendation, traceability and auditability remain too weak for release. If the explanation is contestable and the provenance is preserved, the run may pass. This matters because the empirical paper shows that influence methods as governance interventions shift governance outcomes in different ways: RAG and stacked configurations can improve auditability and traceability when retrieval evidence is preserved, while LoRA/PEFT and alignment layers alter responsibility and dependability profiles in distinct ways. No single method dominates every pillar. Gating therefore matters because it prevents symbolic governance, forces review of the visible profile rather than relying on a reassuring composite alone, and makes remediation actionable when a workflow is not yet ready for deployment or user-facing use.

An HR team uses GenAI to generate a justification for why one applicant was shortlisted and another was not. Under RAIDT, the draft cannot be released to hiring managers simply because it reads clearly. The gate checks whether the run-level evidence pack preserves the criteria version, prompt template, model or adapter identifier, output, and the reviewer check showing that prohibited factors were not used.

If the run lacks those artefacts, the case does not proceed. The organisation may add stricter prompt constraints, bind the workflow to versioned shortlist criteria, or require manual review until the score profile improves. In this way, gating matters because it stops a weakly evidenced recommendation from becoming an organisational decision record.

• 08-RAIDT_Foundations_M_V50
• 09-RAIDT_Empirical_M_V50.docx