What a Development Bank Actually Buys: Explainable-by-Construction

By Anushka Appala and Dr. Janio Rosales

When a multilateral lender evaluates a government technology proposal, it is not, in the end, buying software. It is buying a defensible answer to a single question: when this system acts on a citizen, can the state explain why, and prove a human stood behind it? Everything else — the models, the interfaces, the dashboards — is downstream of that answer. A proposal that cannot give it is a liability dressed as innovation. A proposal that can give it, by construction rather than by promise, is exactly what a responsible-AI framework is built to fund.

This is the design choice at the center of NaciluzIA, and it is worth being precise about what it means.

Explainability as a property, not a feature

Most conversations about "explainable AI" treat explanation as something you add after the fact — a report generated to justify a decision the model already made. That approach inverts the problem. If a probabilistic model produced the outcome, the explanation is a reconstruction, and a reconstruction can be wrong in the same silent ways the model can.

NaciluzIA takes the opposite path. The AI understands, translates, drafts, and flags — but it never approves a benefit, suspends a right, or authorizes a payment. Every binding action is executed by a deterministic, published, versioned rules engine: the same inputs return the same result, every time, for a reason a citizen can read and a court can review. The explanation is not generated after the decision; it is the decision. "Approve if the DPI is valid, the person is living, there is no duplicate benefit, and the socioeconomic threshold is met" is not a model output to be interpreted — it is published law, rendered as code, with a version number.

We call this being explainable by construction. It is the difference between a system that can usually tell you why it acted and a system that is structurally incapable of acting for a reason it cannot state.

Why this is what the framework is looking for

The context matters. Latin America is early in building the legal scaffolding for public-sector AI. Only seven Latin American countries have a national AI strategy — Argentina, Brazil, Chile, Colombia, Mexico, Peru, and Uruguay (OECD-OPSI, Strategic and Responsible Use of AI in the Public Sector of LAC). Guatemala is not among them. Meanwhile, globally, the field is moving fast: the OECD.AI Policy Observatory tracks more than 900 AI-policy initiatives across over 80 jurisdictions (OECD.AI, 2026). The frameworks are proliferating; the deployable, legally grounded implementations are not.

Into this gap, the Inter-American Development Bank published its AI Framework in January 2025 — a responsible-AI standard meant to guide how the institutions it lends to actually build. A framework like that is not looking for the most capable model. It is looking for the property that lets a state deploy AI without surrendering accountability: traceability, human authority over binding acts, and an audit trail that survives contact with a court. A design that is explainable by construction is not a compromise a bank tolerates; it is the specification a bank is trying to fund.

The reason is simple. A framework can require responsible use, but it cannot itself guarantee it. What it needs is architecture that makes irresponsible use impossible — where the guarantee is a property of the system, not a clause in a contract that a future administration might ignore.

The ledger and the law

The second half of explainable-by-construction is proof that the explanation is not tampered with after the fact. In NaciluzIA, every step — the AI's recommendation, the rule version that executed, the official who signed, the timestamp — is recorded in a tamper-evident, hash-chained ledger, valid under Guatemala's electronic-signature law, Decreto 47-2008. Each entry is cryptographically linked to the one before it, so the record cannot be quietly edited without breaking the chain.

This is what gives the design legal standing rather than merely technical elegance. A named human official signs each consequential action, and that signature carries the same weight in Guatemalan law as a signature on paper. The ledger then makes that signature auditable forever. When the Contraloría reviews a decision, it does not ask the system to reconstruct its reasoning; it reads which rule version justified which payment, and who authorized it. The evidence was captured at the moment of the act, not assembled in its defense.

That is a meaningfully different posture than "we log our AI's outputs." It means the state can show, for any decision, the exact deterministic rule that governed it, the human accountable for it, and a record that cannot have been altered since.

An open standard, not a black box

There is one more thing a development bank is wary of, and it is not technical: vendor lock-in. A system that only its builder can inspect is a system the public cannot truly audit, however good its intentions. NaciluzIA is therefore designed as an open, auditable standard — nonprofit-stewarded, with a for-profit delivery arm — and co-built with Guatemalan institutions: UVG, USAC, and the Academia de Lenguas Mayas. The rules engine is published. The ledger format is inspectable. The standard belongs to the public interest, not to a supplier.

This is what closes the loop. Explainable-by-construction is only credible if the construction is visible. An open standard makes the guarantee checkable by anyone — auditors, courts, universities, citizens — rather than asserted by the people who built it.

So when we say a development bank buys explainability, we mean something concrete. Not a promise of good behavior, but an architecture in which the AI recommends, policies decide, officials oversee, and audits verify — and in which each of those is a property you can inspect, version, and prove. That is the product. Everything else is how it gets delivered.

Cada decisión, a la luz.

— Anushka Appala and Dr. Janio Rosales