Computed, never invented

Every analytics tool with “AI” in the pitch will tell you it gives you answers. Almost none of them will tell you where those answers come from. The number on the card — did the model compute it from your rows, or did it write a plausible-looking figure because the sentence needed one? For most tools, the honest reply is that you can’t tell. And if you can’t tell, you can’t trust it.

When blueberry. shows you a total, a trend, or a breakdown, that figure is produced by a query compiler running over your actual rows. The model decides what’s worth measuring — which metric, which grain, which comparison. Deterministic code does the measuring. The chart renders the compiler’s output directly; nothing gets recomputed in the browser, and nothing gets guessed.

That separation is the whole design. The model is remarkable at judgment — noticing that revenue dipped in March, that one region is dragging the average, that two columns are secretly the same thing. It is not where you want arithmetic to happen. So we don’t let it do arithmetic.

Here’s the part most pages leave out. The charts and the KPI cards are structurally computed — those numbers come straight from the compiler. The prose — the summary that says “sales rose after the promotion,” the answer the agent gives you in chat — is written by the model, grounded on those same computed inputs. It’s shaped by strict prompts and read back carefully, but it is a summary, not a second calculation.

What that means in practice: trust the tile. If a sentence and the chart it describes ever seem to disagree, the chart is the source of truth, because the chart is the computed thing. We say this out loud because a tool that overclaims its own accuracy is doing the exact thing we’re trying to fix.

A number you can trust is one you can trace. Ours trace back to your rows, every time.

This is not a modest promise. “Computed, never invented” rules out the failure mode that makes people quietly stop trusting their dashboards — the figure that looked right, got repeated in a meeting, and turned out to be nothing. You don’t reach that kind of trust with a better model. You reach it by refusing to let the model do the part that has to be exact.