The First Caribbean Model Is Also the First One We Can Actually Audit

In late 2025, Maestro AI Labs finished training Maestro. It is a large language model, it was built from the ground up in the Caribbean, and it is the first of its kind in the region. It is currently in red-team testing and has not been released.

CAIRMC has no role in building it, and the Council's Chairman co-founded the lab that did, which is stated here at the top rather than in a footnote. What follows is not an endorsement of the model. It is an assessment of what an original regional model changes for AI assurance in the Caribbean, which is a different question and the only one the Council is qualified to answer.

The Question Assurance Has Not Been Able to Ask

Every AI risk framework in current use, the NIST AI Risk Management Framework, ISO/IEC 42001, the model risk expectations most Caribbean financial regulators have imported from SR 11-7, assumes that somebody can describe how the model was built. Data lineage, development controls, validation evidence, ongoing monitoring. That assumption is load-bearing across all of them.

For a foundation model rented from a foreign provider, the assumption fails at the first line. The Caribbean institution deploying it cannot describe the training corpus, cannot enumerate the fairness evaluations, cannot see the red-team results, and cannot commission an independent test of any of it. The vendor frequently cannot either, because it is reselling somebody else's model.

The practical consequence is that a great deal of Caribbean AI assurance has been performed on the application layer, on prompts, on logging, on human review, while the component carrying most of the risk was recorded as out of scope. That is not negligence on anyone's part. It is the only assurance available when the model is a black box on another continent.

What Changes When the Model Is Built Here

Maestro AI Labs has stated four properties of the model, each of which corresponds to a control that assurance can, in principle, test.

No foreign base model. The parameters were initialised randomly and trained from that point, with no upstream checkpoint. This is not a fine-tune, a LoRA, an adapter, a distillation or a merge. For assurance, this collapses a chain of inherited unknowns into one set of decisions made by a party who can be asked about them.

Document-level data provenance. Training used publicly available data with origin recorded per document. No corpora scraped without permission, no pirated collections. This makes data lineage, the first requirement in every framework listed above, an answerable question for the first time in a Caribbean deployment.

World model methods. The training objective extends past next-token prediction toward holding a consistent internal representation of entities and their states. For risk purposes the relevant consequence is a reduction in intra-document contradiction, which is a specific and testable failure mode in any system summarising legislation, case files or claims history.

Fairness applied during training. Representational balance was part of the optimisation target rather than a post-hoc output filter, evaluated on Caribbean-specific axes: nationality within CARICOM, skin tone, creole versus standard English register, rural versus urban origin, and participation in the informal economy.

That last one deserves an assurance note of its own, because it is where Caribbean deployments fail most often and most invisibly. The disparate-impact harms that matter in this region are usually not statements. They are scores. A model does not announce that an informal earner is unemployed; it ranks her lower, and no output filter has anything to catch. A fairness control that operates on generated text is testing the wrong surface entirely.

Eight Questions Assurance Should Now Be Asking

The existence of a regional model that can answer these is what makes them reasonable to ask of every vendor. A supplier who cannot answer is not disqualified; the answer itself is the finding.

1. Base model lineage. Is this an original model, a fine-tune, an adapter over hosted weights, or a resold API? Name the upstream checkpoint if there is one, and name its licence.
2. Training data provenance. Can the origin of the corpus be described, and by whom? If the provenance belongs to an upstream provider, say so, because that is a material fact for an institution processing citizen data.
3. Fairness implementation. Output filter, post-training adjustment, or training-time constraint? Name the population axes evaluated and attach the results. Distinguish the three rather than using one word for all of them.
4. Evaluation on the deploying population. Not a public benchmark. Results on the institution's own applicants, customers or claim files, with outcomes that can be checked against reality.
5. Red-team scope and findings. What adversarial testing was performed, by whom, against which threat model, and what was found. A report with no findings is a report that was not attempted seriously.
6. Model card availability. Is there a document stating capabilities, known failure modes, evaluation results and the data policy? If not, the institution is being asked to accept an undocumented component into a regulated process.
7. Inference location and control. Where does inference physically execute, and is that written into the contract? A verification step exists and costs an afternoon: run the demonstration on a network with outbound access to the major foreign providers blocked.
8. Continuity if the model is withdrawn. What happens operationally if access ends without notice. This is not hypothetical, and no service-level agreement binds the government sitting above a foreign vendor.

What an Original Model Does Not Fix

Four limitations, stated because an assurance note claiming completeness is worth less than one that does not.

Local does not mean safe. A model trained in the region can be biased, brittle, jailbreakable and wrong in exactly the ways an imported one can. Provenance improves auditability. It does not improve behaviour, and an institution that treats regional origin as a substitute for testing has replaced one unexamined assumption with another.

Capability ceiling. Maestro was trained on Caribbean-scale data and compute. It will not match a frontier laboratory on hard open-ended reasoning, and the lab says so itself. Deploying an under-powered model into a task that needs frontier capability is a risk in its own right, and it is the failure mode the Council expects to see most often as regional models proliferate.

The claims are currently unverified. CAIRMC has not independently tested Maestro and is not in a position to confirm the training methodology. What can be verified, once published, is the model card and the evaluation methodology behind it. The Council's position is that the model card, not the announcement, is the artefact assurance should wait for, and that this standard applies to Maestro on exactly the terms it applies to every vendor.

One model is not an ecosystem. Concentration risk does not disappear because the concentration moved closer to home. A region running critical services on a single regional model has a different dependency, not an absent one.

Why CAIRMC Is Raising This Now

The window for writing these clauses into contracts is while the systems are being procured. Caribbean institutions are buying language-model capability considerably faster than they are building the capacity to assess it, and the assessment gap is widest exactly where the risk is highest: credit, claims, eligibility, triage, and anything that produces a ranking over people.

For three years the honest answer to most of the eight questions above has been that the institution cannot know. That answer is no longer structurally forced. Whether a given vendor can answer is now a fact about that vendor rather than a fact about the state of the technology, and assurance should record it that way.

CAIRMC's related work includes the Caribbean AI risk standard, certification against it, and the TurtleBird safety toolkit built through Maestro AI Labs and made available free to every government in the region for pre-deployment harm testing. The full set of Caribbean AI initiatives this sits within is documented at adriandunkley.net/initiatives.html, and Maestro AI Labs has published its own technical account at maestroailabs.com.