Derived vs. Fetched: Why We Never Blur the Line

Two kinds of numbers

Look closely at anything we show you about a bill and you will find two very different kinds of information. Some of it we fetched — pulled directly from an official source, unchanged. Some of it we derived — produced by a model reading that source and drawing a conclusion.

A vote tally is fetched. The text of a bill is fetched. A cost estimate published by an official scoring office is fetched. These come from a primary source, and we carry them across without editing the substance.

An estimate of how a bill aligns with your values is derived. A summary of what a provision does is derived. A model produced it. It is useful, but it is an interpretation, and it could be wrong.

An estimate dressed up as an official figure is a lie wearing a uniform.

Why the line has to be visible

The danger is not that derived data exists. Derived data is most of what makes the product useful. The danger is when a derived estimate quietly takes on the authority of an official fact — when a model's guess is shown with the same confidence as a recorded vote.

That blurring is how good intentions turn into misinformation. A reader who can't tell the fetched figure from the derived one has no way to calibrate their trust. They might lean hard on a number that was never meant to bear weight.

So we keep the line visible everywhere. Fetched data carries its source and the time we retrieved it. Derived data carries the marks of its making — which model produced it, which version of the method, and whether a human has reviewed it yet. You can always tell what kind of thing you are looking at.

A model's estimate never wears an official's badge

The strictest version of this rule shows up with dollar figures. When an official scoring office publishes a cost estimate, that is a fetched number with a primary source behind it. When our pipeline reads a document and proposes a figure, that is a derived estimate — and it is labeled as one, every time, no matter how confident the model is.

We will not let a derived figure stand in for the official one. If we only have an estimate, we say so. If the official number exists, we show it as the official number, with a link to the document it came from. The two never trade clothes.

Provisional until a human checks

Derived analysis also starts life as provisional. A model's first pass is shown with that status attached, plainly labeled, and the features that make a pointed claim about a person or a policy wait for human review before they go live.

That review is where a derived estimate either earns the right to be treated as reliable or gets corrected. Until then, you see it for exactly what it is: a starting point, not a verdict.

The whole arrangement comes down to one habit. We never let the machine's work impersonate the record. You can read more about how we draw that line in our methodology.