• workshop_dense_100#slide 75

Manual implementation in RAIDT means applying the framework through controlled spreadsheets, saved prompts/outputs, reviewer forms and scoring sheets rather than through integrated logging infrastructure. That low-tech form is fully consistent with the papers because they explicitly state that early pilots may populate run records manually and score them with the rubric. What matters is not automation for its own sake, but whether the organisation can assemble a run-level evidence pack that supports later reconstruction, review, and challenge.

A proper manual implementation therefore still preserves the core RAIDT structure. The organisation scores the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability), produces a score profile for each material run, and uses evidence pointers so that judgements can be checked by another reviewer. The scoring logic remains disciplined through the anchors 1=missing / 3=partial / 5=audit-ready. Manual implementation also requires the team to record influence methods as governance interventions, because prompting choices, retrieval context, adaptation layers, and alignment controls can alter governance outcomes even when the underlying model family is unchanged.

It matters for RAIDT because it lowers the threshold for adoption without lowering the evidentiary standard. Organisations can begin in pilots, training settings, or supervision workshops, learn which artefacts are essential, calibrate reviewers, and identify recurring evidence gaps before investing in automation. In this way, manual implementation functions as a practical bridge from abstract responsible AI commitments to run-level governance practice.

A finance team piloting credit adverse-action explanations can implement RAIDT manually before building any dedicated tooling. For each explanation run, staff save the prompt template, the generated reason statement, the documented lending criteria referenced, and the review notes showing whether reasons were separated from assumptions. They record the run on a scoring sheet and ask a second reviewer to inspect the evidence pack.

This matters because the team may discover that the explanation sounds plausible but cannot be tied back to the exact criteria version used at the time. Manual RAIDT makes that weakness visible early. The pilot can continue under supervision, but the organisation now has a concrete basis for requiring better provenance capture before operational rollout.

• 08-RAIDT_Foundations_M_V50
• 18-RAIDT-Technical-Foundation_M_v04