• workshop_dense_100#slide 41

?D. Evidence Architecture? can be read as the branch of the RAIDT reference wording that sets out how evidence is defined, structured, interpreted, and bounded. In Appendix D, the architecture starts with D.1, which gives the core definition and identifies RAIDT?s two practical outputs: the run-level evidence pack and the score profile. D.2 then explains how the same architecture travels across conceptual, empirical, policy, and sector implementation contexts. D.3 fixes the scope by insisting on the run as the unit of governance. D.4 sets out the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) that the evidence architecture must support.

The branch then becomes operational. D.5 specifies the minimum contents of the run-level evidence pack, including identifiers, prompts, model settings, tools, retrieval snapshots, outputs, integrity checks, and review steps. D.6 explains how those artefacts feed the score profile through anchors 1=missing / 3=partial / 5=audit-ready. D.7 states the boundary conditions: RAIDT supports governance readiness, not a guarantee of correctness, and it depends on lawful, proportionate evidence capture with access controls and retention policies. D.8 adds the implementation logic that treats influence methods as governance interventions, so prompt structure, retrieval, adapters, and alignment controls become governed evidence rather than invisible engineering choices. D.9 closes the branch with worked vignettes showing how the architecture functions in healthcare, public services, finance, HR, and cybersecurity.

A public-service team implementing RAIDT could use the D branch almost as a branch overview for practice. D.1 and D.2 tell them what RAIDT is for; D.3 tells them to define each advisory interaction as a governable run; D.4 tells them what outcomes they must evidence; and D.5 tells them what to log for each case, including prompts, model versions, retrieval snapshots, outputs, and review steps.

They would then use D.6 to turn each evidence pack into a score profile, use D.7 to set access-control and retention limits for sensitive claimant data, and use D.8 to ensure that retrieval and prompt templates are versioned as governed interventions. The vignettes in D.9 provide concrete comparators when calibrating the approach.

• 00-RAIDT_Wording_v2
• 11-RAIDT_Academic_Logic_M_v11