Probabilistic Tools, Deterministic Decisions
By Anushka Appala and Dr. Janio Rosales
There is a category error at the heart of most "government AI" projects, and it is worth naming precisely, because getting it wrong is how a helpful tool becomes an unaccountable one. The error is to treat a language model — a system that produces likely outputs — as if it were a system that produces correct outputs. For a chatbot recommending a restaurant, the difference is trivial. For a state deciding who receives a pension, it is everything.
NaciluzIA is built on the refusal to collapse that distinction. A large language model is a probabilistic instrument. Given a citizen's request in Q'eqchi', it can understand the intent, translate it, summarize the case file, and draft the acuerdo — extraordinary work, and none of it deterministic. Ask the same model the same question twice and it may answer differently. It can be confidently wrong in ways it cannot explain. That is not a defect to be patched away; it is the nature of the tool.
The execution of a public decision must be the opposite. Given the same inputs, it must return the same result every time, for a reason a citizen can read and a court can review. "Approve if the DPI is valid, the person is living, there is no duplicate benefit, and the socioeconomic threshold is met" is not a prediction. It is published law, rendered as deterministic code — a rules engine, versioned and reproducible, that does exactly what the statute says and records exactly why.
The mistake nearly everyone makes is to let the first kind of system do the second kind's job. Our entire architecture exists to keep them apart, and to join them at exactly one point.
The commitment gate
Between any AI suggestion and any real-world effect, we place a single structural chokepoint: the commitment gate. It is not a policy, a review board, or a promise in a procurement document. It is a piece of the system that is physically incapable of producing a binding action unless it holds two things at once — an authorized result from the deterministic rules engine and a verified digital signature from a named human official. Miss either input and nothing happens. The model can propose all day; without the rule result and the signature, the gate does not move.
┌───────────┐
│ CITIZEN │ request, in any of 25 languages
└─────┬─────┘
│ (WhatsApp / voice / web)
▼
┌─────────────────────────────────────────────┐
│ AI ADVISORY LAYER (probabilistic) │
│ understands · translates · drafts · flags │
│ grounded by retrieval (RAG) over cited law │
│ >>> PROPOSES ONLY — no authority to act <<< │
└───────────────────────┬─────────────────────┘
│ proposal + evidence
▼
╔═══════════════════════════════════════════════╗
║ COMMITMENT GATE ║
║ Cannot act without BOTH: ║
║ [1] authorized RULE RESULT ║
║ [2] verified HUMAN SIGNATURE ║
║ Missing either -> no binding action. ║
╚════════╦══════════════════════════════╦═══════╝
│ needs [1] │ needs [2]
▼ ▼
┌──────────────────────┐ ┌────────────────────┐
│ DETERMINISTIC RULES │ │ NAMED OFFICIAL │
│ ENGINE (published, │ │ reviews & signs │
│ versioned law-as-code)│ │ (Decreto 47-2008) │
└──────────┬───────────┘ └─────────┬──────────┘
└───────────────┬────────────┘
▼
┌─────────────────┐
│ REAL ACTION │
│ benefit · permit │
│ · payment │
└─────────────────┘
── TAMPER-EVIDENT, HASH-CHAINED LEDGER ──────────────
records EVERY step above: inputs, model version,
retrieval sources, rule evaluation, official identity,
signature. Nothing reaches the citizen unrecorded.
Read the diagram from the gate outward and the guarantee becomes obvious. The probabilistic layer touches the citizen and touches the draft — it never touches the outcome. The deterministic engine decides what the law requires; the human decides whether to stand behind it; the gate refuses to move without both. There is no path, not even an error path, by which a model's guess becomes a quetzal paid or a right suspended.
Why the separation is the feature
Keeping the two systems apart is not caution for its own sake. It is what makes the whole thing deployable. A deterministic engine is explainable by construction: an auditor at the Contraloría can ask which rule justified which payment and get one answer, always the same one. A signature under Decreto 47-2008 gives the action a human author with legal weight. And the ledger — hash-chained and tamper-evident — means the record of why survives the people and the political cycle that produced it.
This matters most in exactly the place people are most tempted to hand AI the keys: fraud. Our detection stack surfaces patterns and evidence for a human investigator; it does not determine misconduct. It identifies what requires legally authorized human review, and stops there. "The AI caught it" is a claim that dies under cross-examination. "The AI made it discoverable, and a named official acted on it under the law" is a claim you can defend.
The scale of what is at stake justifies the discipline. Corruption drains roughly US$2.6 trillion a year, about 5% of global GDP (UN citing WEF/World Bank, 2018), and Latin America loses an estimated 4.4% of GDP — near US$220 billion annually — to procurement, payroll, and transfer leakage (IDB, Better Spending for Better Lives, 2018). Systems that touch money at that scale do not get to be probabilistic where they bind.
A model that predicts is a gift to an overloaded state. A model that decides is a liability dressed as an upgrade. The commitment gate is how we accept the gift without the liability — and why, in NaciluzIA, the algorithm never gets the last word.
Cada decisión, a la luz.
— Anushka Appala and Dr. Janio Rosales