Beyond a Single Threshold: Uncertainty Source Classification and Type-Stratified Conformal Prediction for Agentic CI/CD

Agentic CI/CD pipelines that produce confidence-gated verdicts tell reviewers that an agent is uncertain — but not why. Two structurally different causes of uncertainty produce identical low confidence scores yet demand different reviewer actions: DATA_UNCERTAINTY (aleatoric — evidence is genuinely ambiguous) calls for closer diff review, while SCOPE_UNCERTAINTY (epistemic — the agent lacks domain knowledge) calls for escalation to a domain specialist. This paper introduces Uncertainty Source Classification: a two-category taxonomy (DATA_UNCERTAINTY, SCOPE_UNCERTAINTY) and a secondary LLM classification prompt that identifies which applies for each flagged consumer verdict in the K11tech Agentic AI QA System. Classification runs concurrently at O(n) LLM calls, adds zero pipeline latency, and never blocks the HITL gate. The two categories exhibit measurably different calibration characteristics, enabling type-stratified conformal prediction thresholds that reduce unnecessary HITL escalation by 23% while preserving coverage guarantees (FNR ≤ 0.10). Evaluation on a 120-PR controlled dataset with 42 classified verdicts yields inter-rater reliability κ = 0.81, confirming the taxonomy is operationally stable. The DATA/SCOPE split is grounded in the established aleatoric/epistemic distinction from Bayesian deep learning. This is Paper 5 in the K11tech Agentic QA series. Source code: https://github.com/K11-Software-Solutions/k11techlab-microservice-qa-system