Q151 - How_do_evidence_architecture_pillar_scoring_and_intervention

Q151 — How do evidence architecture, pillar scoring, and interventions connect?

← RAIDT · Star S7 - Academic Theory and Design Logic · primary item: S7.06 · Mechanisms

Appears in sources
Answer

In RAIDT, evidence architecture connects capture, assessment, and improvement by making the run as the unit of governance. Its core architecture is the pairing of a run-level evidence pack with a score profile, so governance is attached to one configured use rather than to abstract policy claims or model-level documentation alone. The run-level evidence pack preserves the bounded proof object for that use: prompt or instruction, active configuration, retrieved context, output, checks, reviewer actions, identifiers, and provenance. This architecture is what makes the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) inspectable in relation to a specific organisational event.

Pillar scoring then translates that evidence architecture into an explicit judgement of governance readiness. RAIDT uses anchors 1=missing / 3=partial / 5=audit-ready so that claims are grounded in preserved evidence rather than narrative assurance. The score profile does not collapse governance into a single opaque number; it shows where a run is strong, weak, or uneven across the pillars, preserving trade-offs that matter for review, escalation, and learning. In this sense, evidence architecture supplies the reviewable object, and scoring supplies the structured expression of its governance quality.

Interventions connect to both because RAIDT treats influence methods as governance interventions, not merely as technical optimisation choices. Structured prompting, retrieval augmentation, adaptation, and alignment can improve or weaken particular pillars, but only insofar as their configurations are versioned, logged, and attached to the run-level evidence pack. Their governance value therefore becomes observable through changes in the score profile. Mechanistically, the sequence is: evidence architecture makes the run reviewable, pillar scoring makes readiness comparable, and interventions provide the causal levers through which organisations redesign weak runs into stronger governance-ready ones.

Practical example

A healthcare example makes the connection concrete. Suppose a hospital uses GenAI to draft a discharge summary from a synthetic discharge note. In a baseline configuration, staff may retain only the output text, leaving limited evidence about prompt version, retrieval context, or reviewer action. The run-level evidence pack is then thin, so the score profile is likely to show only partial governance readiness, especially for Auditability and Traceability.

If the team introduces a structured prompt that requires an uncertainty section and adds retrieval from an internal clinical guideline corpus, those changes function as influence methods as governance interventions. When the organisation also records the prompt version, model deployment, retrieval snapshot, output hash, and clinician sign-off, the same run becomes reconstructable and challengeable. That richer evidence architecture supports stronger scoring against the anchors 1=missing / 3=partial / 5=audit-ready, particularly because interpretability, auditability, and traceability are no longer inferred from narrative claims but evidenced in the record.

Sources in RAIDT papers
Powered by Forestry.md