The Caribbean Now Has Its Own Model. Fairness Should Stop Being a Claim and Start Being a Requirement

A declaration before anything else. I am President of the Caribbean AI Association and I co-founded Maestro AI Labs, which built the model this article is about. CAIA is an independent association rather than a standards body, nothing here is a rule, and the procurement position I set out applies to Maestro exactly as it applies to every other vendor in the region.

With that said: 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 as far as CAIA has been able to establish it is the first one. It remains in red-team testing.

The association's interest in this is not the announcement. It is what an original regional model does to the range of questions a procurement officer in a small ministry can reasonably ask a vendor without being told the request is unreasonable.

The Word That Keeps Doing Unearned Work

CAIA has written before about sovereign becoming an unverifiable sales claim in Caribbean AI procurement. Fairness is following the same path, and faster, because it is harder to test. A vendor can say a model is fair, a buyer can note that the model is fair, and the file closes. Nobody has lied and nobody has learned anything.

Underneath the word sit three quite different engineering positions. A model with an output filter in front of it. A model adjusted after training to reduce measured disparity on some benchmark. And a model where representational balance was part of the optimisation target while the weights were being learned. These are not degrees of the same thing. They fail in different ways and they cost different amounts, and the region is currently buying all three under one label.

Why a Filter Is the Weakest of the Three

A filter sits in front of a model that has already learned the bias from its corpus. It blocks the most obvious statements and leaves the weights exactly as they were. The bias then leaks through paraphrase, through indirect questions, and through any prompt pattern the filter's authors did not anticipate.

The deeper problem is that the harms that matter most in Caribbean deployments are not statements at all. They are scores. A credit model trained on North American repayment behaviour does not announce that a market vendor in Papine is unemployed. It simply ranks her lower, and the filter has nothing to catch, because nothing was said. The harm becomes visible months later in a portfolio, a complaint file, or a decline rate nobody has disaggregated.

That is the case for putting fairness in the objective. Maestro treated representational balance as part of the training target alongside the language modelling loss, evaluated on axes that matter here and that imported benchmarks do not cover: nationality within CARICOM, skin tone, creole versus standard English register, rural versus urban origin, and participation in the informal economy. It is more expensive than filtering, it constrains the model, and it shows up as lower scores on benchmarks that reward fluent confidence. CAIA considers it the only version of the claim that survives contact with a Caribbean population.

Provenance Is the Second Question

Maestro was trained on publicly available data with provenance recorded at the document level. No corpora scraped without permission, no pirated collections, nothing whose origin cannot be described in writing.

This matters to the association for a reason beyond legal exposure. The Caribbean has spent two years arguing, correctly, that its music, imagery and cultural expression are being absorbed into foreign training runs without consent or compensation. A region making that argument cannot then adopt a regional model built the same way. The first Caribbean model had to be able to answer the question the region has been asking everyone else.

The cost was real. A smaller, cleaner corpus produces a model that knows less. CAIA regards that as the correct trade for a first regional model, and notes that it is a trade the next builder now has a precedent for.

What World Model Training Adds, and What It Does Not

Maestro's training used world model methods: the objective extends past next-token prediction toward holding a consistent internal representation of entities, their states, and how those states change.

The practical gain is consistency. A model with a weak internal representation will place a policy start date in January in one paragraph and March in another, because both are locally plausible continuations. That failure mode is unremarkable in a chat interface and disqualifying in a system summarising legislation or case files for a ministry.

CAIA is not in a position to independently verify the internal claims, and says so. What can be verified, once the model card is published, is the evaluation methodology and the results against it. That is the artefact procurement should wait for, not the launch post.

Two Fields CARICOM Tenders Should Score

CAIA proposed a six-field vendor disclosure schedule for sovereignty claims in August. The existence of Maestro makes two further fields defensible, because a vendor can no longer argue that no model in the region meets them.

• Fairness implementation. State which of the three positions the product occupies: output filter, post-training adjustment, or training-time constraint. Name the population axes evaluated and attach the results. A vendor who has only a filter can say so and compete on other grounds; what should stop is the word standing in for the answer.
• Training data provenance. State whether the origin of the training corpus can be described, and by whom. A vendor reselling a frontier model will often have to answer that the provenance is not theirs to describe, which is a legitimate answer and a material fact for a ministry processing citizen data.

Both are written questions answered before any demonstration, for the same reason CAIA gave in August: a demonstration is designed to persuade, and a written answer is designed to be checked.

What This Does Not Prove

Maestro will not out-perform a frontier laboratory on hard open-ended reasoning, and CAIA would treat any claim that it does as a reason for more scrutiny rather than less. Caribbean institutions should continue to buy frontier capability where it earns the dependency.

What has changed is narrower and, for a procurement officer, more useful. Fairness applied during training and data whose origin can be described are no longer hypothetical standards imported from a policy paper. They exist in a model built in this region, which means the region can ask for them.

Maestro AI Labs has published its own technical account. The wider set of Caribbean AI initiatives this sits inside is documented at adriandunkley.net/initiatives.html.