• workshop_dense_100#slide 62

In RAIDT, scoring anchors are the explicit rubric points that define what counts as weak, partial, and strong evidence when a run-level evidence pack is assessed. They are applied to the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) so that governance readiness is measured through inspectable artefacts rather than through general claims about responsible AI. Put simply, the anchors specify the evidential threshold for saying that a run can be reviewed, contested, and learned from.

Their importance follows from RAIDT's core design choice: the run as the unit of governance. Because generative-AI behaviour is shaped at runtime by prompts, retrieved context, configurations, oversight, and other influence methods as governance interventions, governance must be judged on the evidence preserved for that particular run. Anchors 1=missing / 3=partial / 5=audit-ready make this judgement consistent across reviewers and use cases, while calibration and worked examples reduce drift. They also ensure that the score profile remains diagnostically useful: an organisation can see whether its weakness lies in auditability, traceability, or another pillar and then intervene accordingly. Scoring anchors therefore matter because they convert RAIDT from a set of principles into an operational measurement approach that supports reconstruction, contestability, and improvement.

A finance example shows the point. Suppose a GenAI system helps draft text for an adverse-action explanation. If the organisation can show the prompt template, model deployment identifier, reason-code logic, provenance for the policy or data source used, recorded checks, and reviewer approval, the run can approach the audit-ready end of the anchors.

If, however, the explanation is retained without the underlying provenance or review trail, the output may still appear polished while Interpretability and Traceability remain only partial. The anchors matter because they stop the organisation from confusing fluency with governance quality. Instead, they force attention onto the evidence object that would matter if the decision were disputed by a customer, auditor, or regulator.

• 13-RAIDT-Evidence-Review_M_v10
• 00-RAIDT_Scoring_v1
• 08-RAIDT_Foundations_M_V50