AI Governance &
Algorithmic
Accountability
Research
Isabel Velarde's research is published as practitioner working papers on AI governance, algorithmic accountability, and institutional oversight in Latin America, with a focus on high-informality environments where imported governance models can leave structural conditions specific to the region unaddressed. Her current research program examines institutional asymmetries on both sides of evaluative AI deployment: who can become evaluable at all before assessment occurs, and whether organizations retain sufficient capacity to govern the AI systems they operate once deployment occurs. The work proposes practitioner instruments calibrated for Latin American deployment conditions, including the «Evaluability Boundary Review» («EBR») and the «Institutional AI Dependency Review» («IADR»). Both instruments remain proposed and await empirical validation.
Working Papers
Who the Algorithm Doesn't See
A Practitioner Framework for Evaluative Coverage in Economies with High Informality (Latin America)
Working paper · Version 4, August 2026 · First published May 2026 · Zenodo · All rights reserved ·
This practitioner working paper proposes the «Admissible Evidence Boundary» («AEB») to examine how evaluative systems construct evaluability through the records, signals, and documentary forms they admit as evidence before any score, classification, or eligibility determination is produced. Within a broader evaluative coverage gap it distinguishes «Pre-Assessment Exclusion» («PAE»), where no assessment is produced, from «Restricted Evaluability» («RE»), where a formally valid assessment rests on materially incomplete evidence. The second is the harder to see, because the system keeps producing outputs for everyone it does see.
This paper proposes the «Evaluability Boundary Review» («EBR»), a governance review conducted before deployment at four institutional decision points: vendor disclosure, board oversight, investment due diligence, and public procurement. Five Latin American systems read from the public record serve as illustrations: Sisbén IV (Colombia), RappiCard (Mexico), Plataforma Tecnológica de Intervención Social (Argentina), Registro Social de Hogares (Chile), and SISFOH (Peru). The cases illustrate the framework; the EBR awaits empirical validation. This paper positions evaluability itself as an object of institutional oversight, extending AI governance upstream to the conditions that determine who can be evaluated before algorithmic decision-making begins.
Read on Zenodo → doi.org/10.5281/zenodo.19665794
The Governance Gap Behind AI Deployment
A Documentary Framework for Distinguishing External AI Access from Institutional Capability in High-Informality Economies
Working paper · Version 3, August 2026 · First published May 2026 · Zenodo · All rights reserved · DOI: 10.5281/zenodo.20217978
This practitioner working paper examines the «AI Retained Capability Gap», the divergence between an institution's access to an externally supplied AI-enabled function and the enforceable, proportionate capability it retains to examine, challenge, recover, migrate, replace, or sustain it. Continued delivery can look like possession: an organization may operate a system it has limited capacity to govern.
This paper proposes «Capability Self-Attribution», the mechanism by which continued access to externally supplied outputs may be recorded as evidence that the institution possesses the underlying capability, producing «Unrecognized Provider Dependence». It specifies «Capacity-Output Asymmetry» as the documentary signature of this condition, sets out the «Documentary Coding Protocol» as the method for reading acquisition records, and proposes the «Institutional AI Dependency Review» («IADR»), a two-stage board-level review that surfaces this dependence at acquisition, before deployment. Six documented technology deployments illustrate the framework as structural analogies rather than empirical proof.
The governance problem is general; its regional application is to high-informality economies. The paper synthesizes established reference points, including the NIST AI RMF, ISO/IEC 42001, and the OECD AI Principles, focusing on retained institutional capability rather than system-level properties alone. The constructs and the review remain proposed and await empirical validation.
Read on Zenodo → doi.org/10.5281/zenodo.20217978
Open Access Research
The papers are published on Zenodo, the open-access repository operated by CERN, and released under Creative Commons Attribution 4.0 International (CC BY 4.0). This open-access approach ensures the research remains freely accessible to policymakers, investors, academics, and practitioners across the region, supporting open dialogue and evidence-based governance in contexts where commercial or institutional paywalls would otherwise limit reach. The research is indexed in OpenAIRE, citable through permanent DOIs, and linked to the author's ORCID Identifier (0009-0005-0798-0798).