COMPEL Glossary / uncertainty-quantification
Uncertainty Quantification
Uncertainty quantification encompasses methods for measuring and communicating how confident an AI model is in its predictions.
What this means in practice
Traditional ML models often output a single prediction without indicating whether the model is highly confident or essentially guessing. Uncertainty-aware models provide confidence scores, prediction intervals, or probability distributions that help users calibrate their trust: 'The model predicts 85% probability of churn, with high confidence' versus 'The model predicts 55% probability of churn, with low confidence due to limited data for this customer segment.' Uncertainty quantification is essential for responsible AI deployment because it enables appropriate human oversight -- users can accept high-confidence predictions while applying additional scrutiny to uncertain ones. In COMPEL's agent governance framework, uncertainty thresholds are explicit escalation triggers.
Why it matters
Uncertainty quantification provides the confidence information that enables appropriate human oversight of AI systems. Rather than binary trust-or-reject decisions, uncertainty-aware models help users calibrate their reliance: accepting confident predictions while scrutinizing uncertain ones. This capability is essential for responsible AI deployment in any context where incorrect predictions carry significant consequences for people or the organization.
How COMPEL uses it
In COMPEL's Agent Governance framework, uncertainty thresholds serve as explicit escalation triggers that determine when autonomous agents must defer to human judgment. During the Model stage, confidence scoring requirements are specified for high-stakes systems. The Evaluate stage monitors calibration quality to ensure confidence scores accurately reflect prediction reliability. The Technology pillar assesses uncertainty quantification capabilities as part of model governance.
Related Terms
Other glossary terms mentioned in this entry's definition and context.