There is a particular kind of confidence that AI writing has, and it is unearned. A model will tell you that lobbying spend in a sector rose 41% last quarter in exactly the same tone it uses for things it made up. The prose does not change. The reader has no way to tell the difference.
So we built the difference into the interface instead of the tone.
The mechanism, not the feature
Every tool our agents call returns rows, not summaries. When a model writes a figure into a brief, it marks which rows it used. The interface renders those marks as source chips — small, countable, and clickable. Tap one and you land on the filing, the disclosure, the bill text, or the price series it came from.
That last part is the whole point. A citation you cannot open is decoration. A chip that deep-links into the row is a claim you can audit in about four seconds, which is roughly the amount of patience anyone has for verifying something they were told.
If a figure cannot be traced, it does not ship. Not softened, not footnoted — not shipped.
What this costs us
Plenty. Traceability is the reason some of our outputs take minutes rather than milliseconds: the agent has to actually go and get the material. It is the reason a few things we would like to say on this blog, we cannot. And it constrains the composition of nearly every screen, because chips take room and we will not hide them to win back whitespace.
- Agents return sources or they return nothing.
- Figures carry marks; the UI renders marks as chips.
- Chips deep-link to the row — not to a bibliography.
- AI-generated text is labelled as AI-generated, always.
None of this makes the model smarter. It makes the model checkable, which for money is the more useful property.