• qa_deck_100#slide 63 · Scoring anchors, profiles, and trade-offs

In practice, RAIDT uses anchors 1=missing / 3=partial / 5=audit-ready for each pillar scored from the run-level evidence pack. A score of 1 means that required evidence is absent or the run clearly fails the intent of the pillar. The problem is not simply that the output is poor; rather, the run cannot support accountable review. If prompts, configuration identifiers, retrieval records, oversight decisions, or provenance fields are missing, the run is weakly governable because a later reviewer cannot reliably reconstruct what happened.

A score of 3 means that some important evidence exists, but governance remains incomplete. The papers describe this as partial evidence that may be acceptable only for lower-risk use. For example, a run may retain the prompt and output, yet omit retrieval hashes, key checks, or a clear oversight record. By contrast, a score of 5 means that evidence is complete enough to support reconstruction, review, and justified use for the stated task. In practical RAIDT terms, that means the run-level evidence pack contains the identifiers, logs, checks, and review artefacts needed for an independent audit. The anchor points are deliberately conservative in high-impact settings: citations or fluent explanations are helpful, but they do not justify a high score if the underlying evidence object is incomplete.

An HR performance-appraisal assistant illustrates the difference clearly. If a manager keeps only the final draft appraisal, the run is close to 1 for Auditability and Traceability because the organisation cannot show which prompt template, model settings, policy text, or review action shaped the document.

If the team stores the prompt version, output, and a manager approval note, but not the policy text retrieved or integrity hashes, the run sits around 3: there is some evidence, yet not enough for confident challenge or replay. A 5 requires the full run-level evidence pack: prompt and template identifiers, model deployment details, retrieved policy snapshot where used, output hash, checks, and recorded oversight. The difference is the difference between plausible paperwork and audit-ready governance.

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