• integrated_82#Q2.13

RAIDT is mechanism-based because it treats governance outcomes as generated through recurrent socio-technical processes, not as static properties of a model, a policy, or an ethics statement. The theory paper is explicit that governance in generative AI depends on how one configured use is assembled in context, which is why RAIDT takes the run as the unit of governance. A mechanism-based account is needed because the same model family can produce very different governance conditions depending on prompt structure, retrieval design, adaptation layers, review checkpoints, and evidence retention. In that sense, RAIDT explains governance as something produced through configuration and routine. The run-level evidence pack makes that configuration inspectable, while the score profile renders its consequences visible across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability).

The mechanisms that matter are those that change both behaviour and evidentiary quality. The theory names structured prompting, retrieval augmentation, logging and hashing, parameter-efficient adaptation, preference-based alignment, review steps, and evidence retention practices; the academic-logic paper then treats these as influence methods as governance interventions. Their effects are differentiated. Structured prompting can improve Interpretability and sometimes Responsibility by making reasoning constraints and uncertainty disclosures clearer. Retrieval augmentation can strengthen Traceability and Auditability when retrieval snapshots are preserved. Logging and hashing strengthen reconstructability. Versioned LoRA or PEFT can support Dependability when configuration lineage is preserved. Preference-based alignment can support Responsibility when policy constraints and review signals are explicit. Stacked configurations matter too, because different mechanisms improve different pillars. RAIDT is therefore mechanism-based not for rhetorical effect, but because only this form of explanation can show how governance readiness becomes measurable and improvable through design.

Take a cybersecurity alert-triage workflow. A GenAI assistant helps analysts classify incoming alerts and draft recommended actions. If the team relies on baseline prompting alone, recommendations may vary across runs and later incident review may reveal that no stable record exists of the exact prompt, retrieval source, or review decision.

A RAIDT design would treat influence methods as governance interventions: a structured prompt for consistent triage categories, retrieval from an approved playbook repository, logging and hashing of the run record, a versioned adapter if domain tuning is used, and a human checkpoint for escalation. The run-level evidence pack then preserves the alert context, prompt, retrieval snapshot, output, and reviewer note. The score profile can reveal whether Dependability, Auditability, and Traceability are genuinely strong or only partial. If an alert is later challenged after a missed incident, the organisation can reconstruct the run and ask which mechanism failed: unstable prompting, weak retrieval provenance, missing logs, or inadequate review.

• 12-RAIDT_DSR_Theory_M_v8
• 11-RAIDT_Academic_Logic_M_v11