• qa_deck_100#slide 65 · Scoring anchors, profiles, and trade-offs

RAIDT is trying to preserve trade-offs at two levels. First, it preserves trade-offs across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability), because improvement in one pillar can weaken another if the underlying configuration is poorly governed. The foundations paper is explicit that retrieval grounding may strengthen Traceability while increasing privacy or intellectual-property exposure; alignment layers may strengthen Responsibility while reducing transparency if their policy logic is not documented; and prompt constraints may improve Interpretability while reducing helpfulness when they are overly restrictive. RAIDT therefore frames governance as a multi-objective design problem rather than a search for a single maximised score.

Second, RAIDT preserves trade-offs between evidential strength and the burdens of producing evidence. Because RAIDT treats the run as the unit of governance, it requires a run-level evidence pack that is rich enough to support reconstruction, review, and contestation. Yet the papers also stress that evidence capture has costs: storage overhead, reviewer effort, access-control work, privacy risk, and operational complexity. For that reason, RAIDT does not assume maximal logging in every circumstance. Instead, it aims to make trade-offs visible, reviewable, and proportionate to context.

This is why the framework pays close attention to influence methods as governance interventions. Prompting, retrieval augmentation, parameter-efficient adaptation, and alignment layers are not merely performance tweaks; they change what can be evidenced, how a run can be reconstructed, and which governance risks are created or mitigated. Preserving trade-offs means preserving these differences in the evidence, so supervisors can compare influence configurations on consistent terms rather than on output fluency alone.

In a public-service policy-advice workflow, a team may add retrieval augmentation so the system can ground advice in current policy documents. That choice can improve Traceability, because reviewers can inspect which policy text informed the answer, but only if retrieval snapshots and document versions are stored in the run-level evidence pack. If they are not stored, the organisation gains apparent helpfulness without real reconstructability.

The same configuration also creates a second trade-off. Storing policy snapshots, prompts, and outputs improves Auditability, yet it increases storage cost and may capture sensitive case details. RAIDT preserves that trade-off by requiring supervisors to see both sides at once: stronger provenance and reviewability on one side, and privacy, retention, and operational burdens on the other. The point is not to avoid the intervention, but to govern it explicitly.

• 08-RAIDT_Foundations_M_V50
• 13-RAIDT-Evidence-Review_M_v10