Nexus checks every figure in an AI-written explanation against the figures the model was given. If the reply contains a number that was not supplied, the entire answer is discarded rather than shown.
That is the whole mechanism, and it is unusual enough to be worth stating plainly: the AI in Nexus cannot tell you a number that did not come from your data, because an answer containing one never reaches you.
Why a fluent wrong number is the dangerous case
A language model asked to explain a chart will produce a paragraph that reads correctly whether or not it is. A hallucinated figure does not arrive flagged as uncertain — it arrives in the same confident register as the true ones, in the same sentence structure, surrounded by figures that are correct.
This is worse than an obvious error. An AI that returns nonsense gets ignored. An AI that returns a well-written paragraph with one invented figure in the middle gets quoted in a board deck.
For a platform whose output drives budget decisions, that failure is not acceptable at any rate. The answer cannot be a better prompt or a more capable model, both of which reduce the rate without eliminating it. It has to be structural.
How the check works
The server builds the prompt, including every figure the model is permitted to use. Instructions are owned by the server, not supplied by the browser.
The model writes its explanation.
Every number in the reply is extracted and matched against the figures supplied. Rounding is accepted — prose legitimately says "about 23,000" for 23,041 — but a figure with no source is not.
If any unsourced figure is present, the answer is rejected in full and the user is told the explanation was discarded, rather than shown a partially-verified paragraph.
Rejecting the whole answer rather than the offending sentence is deliberate. A paragraph that contained one invented figure has demonstrated that its reasoning was not anchored, and the remaining sentences carry no more credibility for having avoided a number.
Code computes, the model narrates
The validator works because of a division of labour that runs through the product: arithmetic happens in code, and the model's job is to say what the numbers mean.
Where Nexus decomposes a change in return into its causes, that decomposition is computed and verified to reconcile exactly — the components sum to the observed change, with the residual reported when it does not. Where clusters of campaigns are compared, the patterns are detected by code and the model describes the pattern it was handed.
The model is never asked to calculate. It is asked to explain, which is the thing it is actually good at.
What this rules out
Being honest about the boundary:
It verifies figures, not conclusions. A reply using only supplied numbers can still draw a poor inference from them. The validator catches invented data, not weak reasoning.
It does not make the explanation correct. A correctly-sourced figure interpreted the wrong way passes the check.
It depends on the prompt containing the right figures. The check is only as good as the data assembled for it.
What it guarantees is narrower and more useful than "the AI is accurate": no number you read came from the model's imagination. Every figure in an explanation exists in your data and can be traced back to it.
Why a bolted-together stack cannot do this
This is the structural argument for a single system rather than several.
A copywriting tool connected to a BI dashboard has no access to the computation behind the chart. It receives a rendered image or a flattened export, writes about it, and has nothing to check itself against. The validation step is not something it has chosen to skip — it has no source of truth to validate against.
The check is only possible because the same system computed the figures, assembled the prompt and received the reply. That is what buying analytics, modelling and AI copy separately cannot reproduce, however good each piece is individually.
Model selection
Different tasks get different models. Explanations that inform money decisions run on a frontier model, where the quality of the reasoning is visible to the reader and worth the cost. High-volume classification — labelling thousands of comments, for instance — runs on a smaller, faster model, where the task is narrow and the accuracy gap is small.
Using one model for everything is a mistake in either direction: the expensive model on per-row classification is uneconomic at any real volume, and the cheap one writing a summary produces the flat generic copy that makes a team stop reading the panel.
