An analyst has one afternoon to get the Q3 solar brief in front of the board. The request they type is under-specified, the way real requests are. Steer reads it and splits the gaps in two: what it is willing to assume, and what it refuses to decide for you.
Typed as loosely as anyone actually types it. Every tool accepts this. The question is what it does with the six things the sentence never said.
Two kinds of gap, surfaced separately. Assumptions are shown as values you can change. Forks are questions where a wrong guess wastes the afternoon, so the model leaves them open.
One run guessed the six unstated things. One was told. Marked spans are where the difference shows up in the text.
The contract, as it stands right now. It updates the moment you change an assumption or answer a fork, and nothing is written until it is complete.
Every steer is recorded. When someone asks why the brief says what it says, the answer is a record, not a recollection.
You asked for the 2026 forecast with sources. The draft contains four claims. Ground checks whether the sources exist.
Go to Ground →Press play to watch this run start to finish.
Every loose request has gaps. Most tools fill them silently and hand back something fluent and wrong. This one shows you the guesses and refuses to make the three that are actually yours to make.
About 50 seconds, narrated. Synthetic data. Pause or take over at any point.