• qa_deck_100#slide 11 · Evidence pack and score profile

RAIDT governs the run rather than the model alone because the object that affects organisational action is not a bare model but a configured socio-technical use. The papers are explicit that value and risk in generative AI are shaped by runtime arrangements such as prompt templates, retrieved materials, toolchains, decoding settings, external providers, oversight actions, and adaptation or alignment layers. A model card can describe intended use, evaluation conditions, and known limitations in general terms, but it cannot reconstruct what prompt version, retrieval snapshot, or approval decision shaped one concrete output.

This distinction matters because disputes and reviews usually concern one use event: a credit explanation sent to a customer, a clinical summary entered into a record, or an HR justification relied upon in a contested decision. RAIDT therefore treats run as the unit of governance and operationalises that unit through the run-level evidence pack and an evidence-based score profile. The five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) make it possible to judge whether a particular configured use was governable, while anchors 1=missing / 3=partial / 5=audit-ready prevent vague claims of transparency or responsibility. Governing the run does not replace model-level governance. It complements it by connecting general documentation to actual organisational use, which is the level at which reconstructability, answerability, and contestability become meaningful.

Take the credit adverse-action explanation scenario. Two teams may use the same model, yet one team employs a constrained prompt that links reason statements to documented lending criteria, while another uses a looser template with no preserved rationale structure. If a customer challenges the explanation, the relevant question is not simply which model family was deployed.

The organisation needs the run-level evidence pack for that explanation: the prompt template version, any retrieved policy text, the active model settings, the output hash, and the reviewer's sign-off. That evidence allows the specific explanation to be reconstructed and scored for interpretability and traceability. Model-level documentation alone cannot do that.

• 13-RAIDT-Evidence-Review_M_v10
• 16-RAIDT-Audit-Accountability_M_v05