Assurance Frameworks
Research and evaluation foundations for operational AI assurance.
Ducaltus develops assurance frameworks and evaluation approaches that examine deployment-relevant questions including reliability, fairness, subgroup performance, threshold sensitivity, deployment risk, and governance-aware assessment in high-stakes AI systems.
These frameworks form part of the research and evidence base informing the wider Ducaltus approach to operational AI assurance. The Ducaltus Enterprise platform operationalises this approach by connecting assurance evidence with governance review, remediation, reassessment, reporting, and traceability across the AI deployment lifecycle.
Operational AI Deployment Assurance (OADA)
OADA examines how technical evaluation can be extended toward deployment-oriented assurance for high-stakes AI systems.
The research considers deployment readiness, reliability, fairness, threshold sensitivity, governance requirements, remediation, and reassessment as connected assurance concerns rather than isolated model-evaluation questions.
Fairness Disagreement Index (FDI)
The Fairness Disagreement Index (FDI) examines disagreement between fairness metrics under varying evaluation conditions and decision thresholds.
The research addresses an important limitation in AI fairness evaluation: different fairness metrics can produce conflicting conclusions about the same model, creating uncertainty that may matter when evidence is used to support deployment decisions.
FairRisk-FDI
FairRisk-FDI extends fairness disagreement analysis toward deployment-oriented risk evaluation.
The research examines how fairness disagreement, subgroup variation, threshold behaviour, and deployment context can affect the interpretation of evidence used in high-stakes AI assurance.
Intersectional Fairness Evaluation (IFEM)
IFEM examines intersectional subgroup performance and disparities that can remain hidden by aggregate accuracy or single-attribute analysis.
The research considers subgroup reliability, threshold-sensitive behaviour, mitigation outcomes, and variation across combined demographic conditions as part of deployment-relevant evaluation.
Subgroup Reliability Evaluation
Ducaltus evaluation approaches emphasise subgroup-level evidence alongside aggregate model performance.
This includes consideration of error behaviour, performance variation, threshold sensitivity, and disparities across relevant demographic or operational groups where supported by the available evidence.
Deployment-Oriented Evaluation
High-stakes AI assurance requires consideration of how technical evidence relates to the intended deployment context rather than treating predictive performance as the complete deployment decision.
Ducaltus therefore considers reliability, uncertainty, subgroup behaviour, error trade-offs, operational conditions, and governance requirements as part of deployment-oriented assurance.
From Research to Operations
Operationalised through the Ducaltus platform.
Research and evaluation frameworks provide part of the evidence base. Ducaltus Enterprise provides the operational assurance environment around that evidence, supporting structured review, governance oversight, remediation, reassessment, reporting, and traceable assurance activity across the deployment lifecycle.
This separates the underlying research and evaluation foundations from the operational platform used to support repeatable organisational assurance.