Purpose & accountable owner
Document intended task, affected users, business owner and prohibited uses.
Responsible AI assurance
AI Assurance, Evaluation & Impact Assessment
A reusable control framework for evaluating AI systems before and after deployment. It extends Raeburn's existing AI transparency, model cards, responsible-use policy and governance reference architecture.
Data & privacy
Classify inputs/outputs, personal/confidential data, retention and provider boundaries.
Accuracy & task quality
Use task-specific test sets, error taxonomy, citation/grounding checks and acceptance thresholds.
Security
Test prompt injection, untrusted content, data exfiltration, excessive agency, tool misuse and permission boundaries.
Autonomy & consequence
Record tools/actions, reversibility, financial/external consequence, approval gates, spend/iteration limits and kill path.
Fairness & impact
Assess affected groups, accessibility, contestability and plausible adverse impacts proportionate to use.
Reliability & resilience
Measure failure rate, fallback, timeout, provider dependency, observability and incident response.
Cost & performance
Track latency, token/compute cost and quality/cost trade-offs rather than benchmark quality alone.
Human oversight
Define when review is mandatory, reviewer competence, escalation and who remains accountable.
Change management
Re-evaluate material model, prompt, tool, data-source and permission changes; retain version/evidence history.
Assurance status
Inventory → Impact screen → Evaluation plan → Tested → Approved with controls → Monitored → Re-evaluated/retired. A framework assessment is not a statutory certification or guarantee of safety.