Typed Indeterminacy Engineering: A Practitioner's Design Pattern for Carrying Uncertainty Types Through a Decision Pipeline

Resumo

Decision systems that operate under uncertainty — LLM-as-judge evaluators, multi-criteria ranking procedures, evidence-fusion pipelines, AI alignment mechanisms, and expert-consensus methods — routinely reduce heterogeneous epistemic failure modes to a single scalar of “confidence.” Established literatures resist this collapse in pieces, including Dempster–Shafer theory, the aleatoric/epistemic distinction, neutrosophic sets, imprecise probability, belief-change theory, selective prediction, conformal prediction, and learning to defer. This article states a design pattern connecting them: a 2 × 3 type grid—failure mode (ignorance, conflict, falsity) × removability (removable, constitutive)—is indexed to the applicable belief-change operation and then to the appropriate decision-output form. We call this pattern Typed Indeterminacy Engineering (TIE). Two worked examples from the author’s applied practice and an application to Delphi-style consensus methods illustrate the pattern, while the article clearly distinguishes conceptual contribution from the formalization and independent validation still required.