• qa_deck_100#slide 62 · Scoring anchors, profiles, and trade-offs

RAIDT uses scoring anchors because it treats the run as the unit of governance and therefore needs a repeatable way to judge whether a specific use is governable in evidence terms. The papers argue that governance claims often remain too abstract when they rely on model cards, policy statements, or narrative assurance alone. A scored run-level evidence pack makes the relevant question concrete: can an independent reviewer reconstruct what happened, inspect the configuration, understand the oversight that occurred, and challenge the basis for reliance? The anchors turn these questions into explicit thresholds rather than leaving them to impressionistic judgement.

Anchors also matter because RAIDT is designed around the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability), and these pillars must be comparable across runs, reviewers, and sectors. By specifying anchors 1=missing / 3=partial / 5=audit-ready, RAIDT separates governance readiness from task fluency or rhetorical quality. This is particularly important because influence methods as governance interventions can improve one pillar while weakening another: retrieval may strengthen grounding, for example, while poor snapshotting weakens traceability. Anchored scoring therefore produces a score profile that keeps trade-offs visible and identifies concrete remediation actions, such as better logging, stronger provenance capture, or stricter human review. In this sense, scoring anchors are not an add-on to RAIDT; they are the mechanism that makes run-level governance inspectable, contestable, and operationally useful.

Consider a public-service policy advice run. The generated answer may look sensible and may even cite a policy rule, yet the organisation has not stored the retrieval snapshot or the exact policy version consulted. Without scoring anchors, reviewers may still be tempted to accept the run because the prose appears authoritative.

Under RAIDT, the anchors force a different judgement. Auditability and Traceability remain low because the run-level evidence pack cannot independently reconstruct which source text informed the advice. The weak score profile then triggers a practical intervention: preserve the retrieval snapshot, version identifiers, and reviewer sign-off for future runs. The value of anchors is therefore procedural and corrective, not merely descriptive.

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