When Missing Evidence Becomes False Certainty: Testing Decision Sufficiency Under Partial Observability Across AI-Assisted Systems

Artificial-intelligence systems are increasingly asked to make consequential decisions from incomplete records, changing environments, selective observations, and outputs whose confidence may be better characterized than their evidentiary foundations. A missing measurement, an observation with negative results, and evidence supporting the absence of a condition are not interchangeable. This paper develops the Epistemic Completion Boundary Hypothesis (ECBH): some unjustified conclusive decisions may arise when a system implicitly treats its available representation as sufficiently complete for a particular decision, although task-relevant observational or inferential conditions remain unmet. A Field-oriented Evidence Sufficiency Contract is proposed as a candidate relational representation of observation coverage, missingness, competing hypotheses, decision requirements, loss, temporal validity, provenance, and authority. The proposed representation is not asserted to be a new physical field, a universal ontology, or an established improvement over Bayesian decision theory, partially observable Markov decision processes, selective prediction, conformal methods, or typed provenance graphs. Instead, the paper specifies observational-equivalence conditions, strong evidence-matched comparison arms, falsifiable hypotheses, independent labeling rules, and a prospective cross-domain study spanning autonomous commerce, indoor safety mapping, and clinical decision support. The primary proposed outcome is unjustified conclusive decision rate, evaluated jointly with valid decision coverage, selective risk, calibration, value of additional information, and operational cost. No experiment, human study, clinical validation, or comparative performance result is reported. The intended contribution is a disciplined research protocol for determining whether explicit relational qualification adds measurable value or is reducible to competent conventional engineering.