• integrated_82#Q2.8

In RAIDT, contestability means that an output, recommendation, or downstream decision can be challenged by a stakeholder because the evidence for that specific use is inspectable rather than merely asserted. Reconstructability is the companion property: the organisation can later recreate what happened in one configured use by treating the run as the unit of governance and preserving a run-level evidence pack that records intent, configuration, provenance, outputs, and oversight actions. Together, these ideas move governance from policy statements and model-level documentation to inspectable proof about one run in context.

This matters because generative AI risk is materially shaped at run time. The same model can produce different governance risks depending on prompts, retrieved sources, enabled tools, settings, and human review. RAIDT therefore uses the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) and a score profile to make governance readiness visible for each run. In that logic, contestability is not only a right to disagree; it is an operational capability supported by evidence. Reconstructability is not simple logging; it is the ability to review, compare, and examine how an output came to be. When evidence is weak, the score profile should show that weakness through anchors 1=missing / 3=partial / 5=audit-ready. These properties matter because they support oversight, exception handling, audit sampling, learning across runs, and more credible accountability when outputs are questioned. They do not guarantee substantive correctness, but they make governance claims reviewable in practice.

Consider an infrastructure firm using generative AI to support a maintenance-prioritisation decision under uncertain conditions. A manager receives a recommendation to defer inspection on a high-cost asset, but an engineer contests the output because recent site notes suggest rising risk. Under RAIDT, the organisation does not rely on a generic policy or a model card alone. It reviews the run-level evidence pack for that specific use: the prompt version, retrieved inspection records, active tools, model settings, generated output, and the human checks applied before the recommendation was circulated.

That reconstruction allows the team to see whether the recommendation rested on incomplete retrieval, weak provenance, or inadequate review. The score profile across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) then shows whether the run was governance-ready. If traceability of retrieved sources is only anchors 1=missing / 3=partial / 5=audit-ready at a partial level, the recommendation can be paused, re-run, or escalated. Contestability therefore becomes a practical review path, not a rhetorical principle.

• 01-Responsible_AI_for_Managerial_Decision-Making_Under_Uncertainty-V3
• 13-RAIDT-Evidence-Review_M_v10
• 15-RAIDT-IS-Governance_M_v07