Computed, never guessed.

Language models are persuasive with numbers and unreliable at producing them. They do not calculate; they perform calculation, the way an actor performs surgery. Most of the time the performance is close. Close is worthless in operations.

So our systems run one strict rule: every figure is computed in code. Revenue, reach, feasibility, anomalies — each is the output of a query or a program that can be read, tested, and rerun. The language model's job begins after the number exists: it narrates. It says what moved, what looks wrong, what deserves your attention. It is never the source of the value, only the explanation of it.

The same discipline shapes how numbers are shown. Every figure on our pages carries an as-of date, because a true number without its time is halfway to a false one. Where a claim is a client's to make, it is named on permission or kept anonymous — not decorated.

This split — computation for facts, language for meaning — sounds obvious written down. It is surprisingly rare in practice, because the shortcut is so tempting: the model will happily give you the number and the story in one breath. The story will be excellent. The number will be an impression.

We think the businesses that win with AI will be the ones that keep this line bright: machines that are honest about which of their outputs are facts and which are judgment, presented to a human who owns the final call. It is how our own operations run, every day, three ventures at a time.