• qa_deck_100#slide 61 · Interpretability, dependability, and traceability

RAIDT keeps interpretability, dependability, and traceability separate because each pillar expresses a different governance claim about a run-level evidence pack. Interpretability asks whether stakeholders can understand, evaluate, and appropriately rely on an output in context. Dependability asks whether the system behaves stably and reliably under expected conditions of use, including repeated runs and minor prompt variation. Traceability asks whether reviewers can establish provenance: what sources, transformations, models, tool calls, adapters, and policy layers contributed to the output. These are not interchangeable qualities, and the papers treat them as distinct outcomes within the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability).

The separation is methodological as well as conceptual. RAIDT uses anchors 1=missing / 3=partial / 5=audit-ready and scores the run as the unit of governance, not the model in the abstract. That means each pillar must be justified by different evidence fields in the run-level evidence pack: explanation templates and uncertainty disclosures for interpretability, repeat-run outputs and dispersion measures for dependability, and retrieval snapshots, identifiers, versions, and hashes for traceability. A single merged score would allow one kind of evidence to compensate for the absence of another, even when the missing evidence blocks reconstruction or contestability.

Keeping the pillars separate also supports action. Structured prompting may improve interpretability without improving stability; retrieval grounding may improve traceability without making the response more comprehensible. Because RAIDT treats influence methods as governance interventions, the profile must show which intervention strengthened which governance outcome, and where residual weakness remains.

In finance, consider a GenAI system drafting a credit adverse-action explanation. With a constrained template and clear uncertainty wording, interpretability could score well. If the same case is run three times and the ordering or wording of reasons shifts materially, dependability remains weaker. If the bank cannot preserve the retrieved policy text, prompt template version, and model deployment identifier, traceability is weaker still.

A composite near the middle might look unremarkable, but the separate profile tells a reviewer something decisive: staff may understand the explanation, yet the organisation still cannot show consistently why that explanation appeared or reconstruct the exact evidence chain months later. The appropriate response is therefore targeted logging, snapshotting, and repeat-run control, not a vague decision to improve the model.

• 08-RAIDT_Foundations_M_V50
• 00-RAIDT_Scoring_v1