An AI system can pass its fairness, explainability, and robustness checks, report a fully healthy dashboard, and still never see a large share of the people it was built to serve. The reason sits upstream of every metric a board currently watches: before a system scores anyone, what it will accept as evidence has already decided who it can evaluate at all. Whoever cannot produce that evidence falls outside the system before any score, ranking, or eligibility decision is made. This paper names that condition «Pre-Evaluative Exclusion». It proposes the «Admissible Evidence Boundary» as the mechanism that produces it, and defines 2 failure modes. In «Absent Evaluability» the system openly fails to return a result, and the gap is visible. In «Restricted Evaluability» it returns a confident result built on a materially incomplete record, and the gap is silent: the institution's dashboards show a system working correctly while it never sees part of the population it is meant to serve. The claim at the center of the paper follows: an institution can believe it is evaluating its whole population while it is only evaluating the subset able to generate admissible evidence, and it can hold that belief precisely while its systems report success. In Latin America this is not a marginal case. Where roughly 47% of workers, some ~150 million people, earn a living outside formal documentary systems, the excluded are a structural majority rather than an edge case. The systems most exposed are the ones meant to reach them: social benefits, credit, and public administration, where technically sound platforms can still miss the people they were built to serve. This paper examines 5 such systems across the region: SISBEN IV (Colombia), RappiCard (Mexico), PTIS (Argentina), Registro Social de Hogares (Chile), and SISFOH (Peru). To act on it, this paper proposes the «Pre-Evaluative Audit» (PEA), an instrument that needs no access to source code or training data and runs at 4 decision points leaders already control: vendor disclosure, board oversight, investment due diligence, and public procurement. It turns a coverage gap that today has no name, no owner, and no document into a finding a board can act on, surfaced before deployment, when correction is cheapest. In these economies, «Pre-Evaluative Exclusion» is almost always present, and it hides in plain sight. The institutions most certain their systems reach everyone are usually the ones that took the dashboard at its word. This paper gives them the way to look behind it, while the gap is still theirs to correct.

This version updates the DOI displayed on the cover page and suggested citation to the concept DOI (10.5281/zenodo.19665794)