• qa_deck_100#slide 68 · Prompting as a governance intervention

In RAIDT, prompting is assessed through a common evaluation protocol rather than described impressionistically. The prompt paper formalises zero-shot, instructional, and role-based prompting as pre-specified patterns, each stored as a versioned artefact with a stable identifier. These patterns are then tested through controlled runs, with fixed configurations, logged metadata, and reviewer scrutiny across domains. The aim is not to ask whether a prompt sounds sensible, but whether it produces measurable governance effects. Assessment is organised around the five RAIDT dimensions and synthesised into a score profile, usually expressed as a radar-style profile that shows where a prompt family is strong and where it remains weak.

The scoring method is explicit. Two reviewers score each output, give short justifications, and trigger adjudication when disagreements exceed one point. The papers also specify concrete pillar anchors, which can be summarised as anchors 1=missing / 3=partial / 5=audit-ready, with dimension-specific definitions such as missing versus full logs, opaque versus audience-fit explanations, and absent versus complete provenance. Because the run as the unit of governance is preserved in the logs, each judgement can be read against a run-level evidence pack containing prompt ID/version, model metadata, hashes, and related artefacts. Prompting is therefore assessed as an evidentially tractable intervention: it is scored, challenged, replayed, and compared, not merely narrated.

Consider consumer-credit explanations. A team may compare a zero-shot prompt with an instructional prompt that requires a clear adverse-action explanation, one actionable counterfactual, and an uncertainty statement. Reviewers do not simply note that the second output feels better written. They score both variants on Responsibility, Auditability, Interpretability, Dependability, and Traceability.

If the instructional prompt consistently produces clearer explanations with fewer omissions, it will receive a stronger score profile on Interpretability and Responsibility. If the run logs are complete, Auditability rises too. However, if the explanation still lacks source grounding, Traceability remains partial rather than audit-ready. That assessment is attached to the logged run, so the organisation can show why a prompt was retained, revised, or rejected.

• 04-RAIDT_Prompt_Eng_V2
• 05-RAIDT_LoRA_V2