Who the Algorithm Doesn't See: A Practitioner Framework for Evaluative Coverage in Economies with High Informality (Latin America)

Abstract

An evaluative system can pass its fairness, explainability, robustness, and operational checks, continue producing outputs, and still fail to assess a meaningful share of the population it is intended to serve. The reason sits upstream of the metrics institutions usually monitor: before a system produces a score, classification, recommendation, or eligibility decision, its evidence requirements have already determined whom it can evaluate at all. The resulting divergence between the population an institution intends to assess and the population its systems can effectively evaluate is the evaluative coverage gap examined in this paper. The paper proposes the Admissible Evidence Boundary (AEB): the evidence requirements, data sources, access channels, documentation rules, technical configurations, and operational practices that jointly determine which records can generate an evaluation. Within the broader evaluative coverage gap, it distinguishes two conditions. In Pre-Assessment Exclusion (PAE), a person, household, or entity receives no assessment and no evaluative output. In Restricted Evaluability (RE), an assessment exists, but the admitted evidence represents the subject in a materially incomplete manner. An institution can therefore mistake the population its systems can evaluate for the population it intends to reach, even while those systems continue operating as specified. In Latin America, this is a structural governance concern rather than a marginal technical exception. The International Labour Organization estimated that 46.7% of employed persons in the region were engaged in informal employment during the first half of 2025. In contexts of high labor informality, uneven financial inclusion, and fragmented administrative records, the absence of compatible documentation does not necessarily indicate the absence of income, productive activity, repayment capacity, vulnerability, or need. The paper examines five documented regional systems as analytical illustrations: Sisbén IV in Colombia, the RappiCard credit scoring model in Mexico, the historical Plataforma Tecnológica de Intervención Social (PTIS) in Argentina, the Registro Social de Hogares in Chile, and SISFOH in Peru. These cases illustrate the framework; they do not validate it. To make this divergence reviewable before deployment, the paper presents the Evaluability Boundary Review (EBR), a proposed governance review that examines the relationship between intended and effectively evaluable populations at four decision points institutions already control: vendor disclosure, board oversight, investment due diligence, and public procurement. The EBR asks whom a system can actually recognize through the evidence it admits, what relevant populations may remain outside that boundary, and whether its documentary and calibration assumptions fit the intended deployment context. The EBR has not yet undergone empirical validation; the paper specifies the validation agenda through which that status can change. Its purpose is to turn an often-unexamined coverage exposure into a reviewable institutional decision while correction remains possible.