
A model cannot tell a verified figure from an unverified one
Calaa is the group's intelligence layer. What decides its quality is the corpus underneath it; the model is the part that sets no one apart. Mind verifies that corpus before it goes in.
Why verification comes before the model
Producing a report with artificial intelligence now takes ten minutes. Answering where each of its figures comes from costs what it always cost, and that is the work Mind does.
A model trained on figures without provenance produces answers just as fluent as one trained on verified figures. The difference does not show when you read them. It shows six months later, when someone makes a decision on one of them.
The two answers
Both read the same. Only one can be checked
Two answers to the same brief, one over a verified corpus and one over an unverified corpus. What changes is not in the text, it is in what each figure carries underneath. The figure compares structure, not wording.
Where does this figure come from?
The difference does not surface as you read. It surfaces when someone asks where a figure comes from, and by then the decision is already made.
The three rules a figure passes before it goes in
They apply to each figure, not to the finished report.
- 01
Reliable sources
Official institutions, public bodies, academic studies, specialist publications and established research firms. When the only source available falls short of that standard, the figure goes in flagged or it does not go in.
- 02
Cross-checking
Every figure is checked against several sources to rule out bias. When two reliable sources disagree, both are kept instead of picking the convenient one.
- 03
Critical review
The source, the date, the method and any interest behind it are assessed. A correct figure in the wrong context misleads just as well, and it is the most common way to get something wrong with a number that is true.
The three gates
None of the three rules answers yes or no
The path a figure takes through the three rules, in the order they apply. Each gate has its own exit. It orders the path, not how many figures each one takes.
Drag to turn the model, or press a gate to see where what does not pass comes out.
All three apply while the figure is going in, not when it is published. A figure that reaches the end carries the label of where it passed, and that label travels with it into the corpus.

01.1Before the model
A model does not ask where a figure comes from. Someone has to have asked first.
Whoever verifies decides what goes in, and that decision is written beside the figure.
What gets flagged, and why the model needs it flagged
Three labels that travel with the figure and survive aggregation.
This is an estimate
An estimate presented as a measurement spreads without warning.
When the figure is a calculation of our own and not a measurement, it is stated and how it was arrived at is explained. After that it cannot be corrected, because whoever received it used it as though it were an observed figure.
This cannot be answered
A model always returns an answer, including when the data does not settle it.
A question the available data does not settle is flagged as open, with what it would take to settle it. A model always finds something to say, so the gap has to be written down by someone first.
This comes from here
The flag travels with the figure and does not end up in an annex.
Every figure keeps its source, its date and the criterion it was accepted on, beside the figure. That way it is still there when someone carries it into another document, which is when it usually gets lost.
What Calaa returns
The intelligence layer aggregates, connects and answers over a corpus Mind has verified. When an answer rests on a figure flagged as an estimate, the flag is still there. That is the only way the group can use its own intelligence to decide and not merely to draft.
The path of the flag
The flag is still there when the answer comes out the other side
The four jumps a figure makes from the moment it is verified to the moment it supports an answer. The figure follows one flag; it does not count how many figures pass through each layer.
The flag
This is where it is lost if it does not travel declared
Without the flag, the answer reads just as well and cannot be checked. With it, the group decides with its own intelligence instead of only drafting with it.

