• workshop_dense_100#slide 17

In RAIDT, an evidence pack is the structured, reviewable record of one run. At minimum, it links the run identifier, timestamp, task and domain labels, initiating role, full prompts and template versions, model and decoding settings, enabled tools, retrieved context where applicable, outputs, integrity hashes, review actions, and governance metadata such as access control and retention. The papers treat this package as a proof object rather than as descriptive prose. Its purpose is to preserve what actually happened so that the organisation can later reconstruct, challenge, and improve a material use of GenAI.

Why it matters is equally central. RAIDT does not claim that an output is correct simply because it exists; instead, it asks whether the evidence is sufficient to justify reliance and oversight. The run-level evidence pack is therefore what makes the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) assessable in a disciplined way and what underpins the score profile. It also carries the practical burden of governance integration: secure storage, access control, retention rules, and linkage to incident tickets, audits, or risk registers. Because RAIDT treats influence methods as governance interventions, the evidence pack must also preserve prompting, retrieval, adapter, or alignment choices where they shape behaviour. Without the pack, governance collapses back into post hoc narrative; with it, the anchors 1=missing / 3=partial / 5=audit-ready can be applied to inspectable artefacts.

A clear example appears in healthcare. In a chest-pain note-summarisation run, the evidence pack would record the fixed clinical schema, the prompt version, the model version, any retrieval used, the output hash, the clinician oversight rule, and the retention tag attached to that run. If early testing showed that explanations were uneven for non-specialist readers, the team could revise the explanation template, repeat the runs, and compare the new evidence packs with the earlier ones.

The importance is practical as well as conceptual: auditors and regulators can inspect a time-stamped chain of prompts, configurations, outputs, and reviewer decisions instead of accepting a static assurance statement. That is why the evidence pack matters in RAIDT.

• 08-RAIDT_Foundations_M_V50
• 11-RAIDT_Academic_Logic_M_v11