A presidential run here puts 51 state delegations in front of two platforms, has them argue, and reports what they concluded. It reads polling data at no point. It has no prior, no house effects file, no turnout model and no error bars, because it has nothing to put error bars around.
So it is not a forecast, and every page that shows a result says so.
The distinction that actually matters
A forecast is a claim about the world: 63% means that if you ran this situation a hundred times, the outcome happens 63 of them, and you can score it afterwards against what occurred. A simulation is a claim about a model: this is what these agents, reading these arguments, under these rules, concluded. You can check whether the model did what it says it does. You cannot check it against November, because it was never a statement about November.
Conflating the two is the standard failure of AI election tools, and it is usually done through the interface rather than the copy. A number rendered as a percentage on a map of the United States reads as a forecast no matter what the paragraph underneath it says.
What it is useful for
The question this answers is not who wins. It is how an argument between two platforms plays out differently in Vermont and in Wyoming, given what those places are actually like. The interesting output is the deliberation text, not the vote share, and the vote share exists mainly to make the deliberation comparable across 51 places.
That is a smaller claim than a forecast. It is also one this site can actually support.
The parts that are not simulated
Verifiable facts are kept out of the model entirely. Election dates and offices come from public records. District geography is computed from Census boundary files, with the method and its failures written up. Headlines are linked to the publisher that wrote them and are never rewritten. Where the AI produces something, it is labelled as a simulation at that point on the page and not only in a footer.
The line is drawn where it is because a model that hallucinates a county name is a bug that can be caught, and a model that hallucinates a plausible-sounding vote margin is a bug that cannot. One of those has already happened here, and it is written up on the methodology page rather than quietly fixed.