Somebody Finally Wrote Down Where the Training Data Came From

Every data governance framework worth the name starts in the same place. Where did this data come from, who owns it, what are we permitted to do with it, and can we show our working.

Then large language models arrived, and an entire generation of data teams quietly suspended that discipline for the single most consequential component in the stack. Ask a vendor what their foundation model was trained on. You will get a paragraph about diverse, publicly available sources, and then the conversation moves on, because there is nowhere for it to go. The vendor frequently does not know either, since it is reselling somebody else's model.

In late 2025, Maestro AI Labs finished training Maestro. It is the first large language model built from the ground up in the Caribbean. It is still in red-team testing. The part of it that should interest a data team is not the capability. It is that somebody wrote down where the data came from.

What "provenance at the document level" means

Maestro was trained on publicly available data, with the origin of each document recorded. No corpora scraped without permission. No pirated collections. No material acquired through a chain that nobody can describe.

That is the same standard a Caribbean bank applies to a third-party data feed before it goes anywhere near a credit decision, and the same standard a market research team applies to a sample frame. It is unremarkable practice everywhere except in foundation models, where it has been treated as impossible.

It is not impossible. It is expensive, and it produces a smaller corpus, and therefore a model that knows less. The lab took that trade deliberately. For anyone who has sat in a model validation meeting, the calculation is familiar: a component whose lineage cannot be documented is a component that fails review, regardless of how well it performs.

Fairness that operates where the harm actually is

The second methodological choice matters more to analytics than to almost anyone else.

Most deployed fairness work is an output filter. A model learns the bias from its corpus, and a second system blocks the worst statements. Maestro instead applied fairness constraints during training: representational balance as part of the optimisation target, evaluated on axes that exist here and not in imported benchmarks. Nationality within CARICOM. Skin tone. Creole versus standard English register. Rural versus urban origin. Participation in the informal economy.

Here is why a filter is close to useless for our work. The disparate-impact harm in an analytics pipeline is almost never a sentence. It is a number. A model does not write that a higgler in Coronation Market is unemployed. It returns a lower score, the score feeds a ranking, the ranking feeds a decline, and the decline appears in a portfolio six months later as a clean, defensible, entirely wrong pattern. There is no offensive output for a filter to intercept. There never was.

Anyone who has run a disparate-impact analysis on a Caribbean portfolio has seen this. The model is not rude about informal earners. It is simply confident that they are riskier, because the data it learned from came from an economy where informal income is rare and usually a red flag. In ours it is normal and frequently a sign of exactly the resourcefulness a lender should want.

World model methods, and one specific failure they reduce

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

The failure mode this addresses will be familiar to anyone who has tried to use a language model over a long document. Ask for a summary of a forty-page policy and the model tells you the effective date is January in one paragraph and March in another, because both continuations were locally plausible and nothing forced it to remember its own earlier commitment. That is a nuisance in a chat window and a defect in a system summarising claims history, loan files or regulatory submissions.

StarApple Analytics has not independently benchmarked this, and will not claim a result it has not measured. What can be said is that intra-document consistency is measurable, and that a model card is due at release.

What this does not mean for your stack

Maestro was trained with Caribbean-scale data and compute. It will not beat a frontier laboratory on hard open-ended reasoning, and nobody involved is claiming it will. A regional model deployed into a task that genuinely needs frontier capability is its own risk, and it is the mistake to expect as regional models multiply.

The sensible architecture for most Caribbean data teams is unchanged: a model you control as the always-on baseline, with frontier capability called for the narrow set of tasks that earn the dependency, so that losing the frontier degrades your pipeline instead of stopping it.

Nor does regional origin make a model safe. A model trained here can be biased, brittle and wrong in exactly the ways an imported one can. Provenance improves auditability, not behaviour, and a team that treats local origin as a substitute for testing on its own population has swapped one unexamined assumption for another.

Three questions to put to every model vendor now

• Can you describe the training data provenance, and if not, who can? A vendor reselling a foreign model will often have to answer that the provenance is not theirs to describe. That is a legitimate answer and a material fact for anyone processing customer data.
• Is fairness a filter, a post-training adjustment, or a training-time constraint? Three different engineering positions are being sold under one word. Ask which one, then ask for the evaluation axes and the results.
• What are the results on our population, not a public benchmark? Your applicants, your customers, your claim files, with outcomes that can be checked. This is where a model trained elsewhere reveals what it assumes about people here.

None of these require a technical evaluation. They are written questions, they are answerable, and for three years the honest reply to the first one was that nobody could know. That is no longer structurally true, which means an unanswered question is now a fact about the vendor rather than a fact about the technology.

StarApple Analytics, part of StarApple AI, works with Caribbean organisations on data governance for AI systems: lineage documentation, disparate-impact testing on local populations, and evaluation sets built from your own data. Maestro AI Labs has published its own technical account, and the wider set of Caribbean AI initiatives this belongs to is documented at adriandunkley.net/initiatives.html.