What it does
BOARDROOM puts a research question to several models with different training and different failure modes, then keeps the transcript of where they disagreed rather than collapsing it into one answer.
A single model asked a hard question returns a confident paragraph, and the confidence carries no information about whether the paragraph is right. Running the same question across models makes the shape of the uncertainty visible: the parts everyone agrees on, and the parts that move depending on who is asked.
In the code
@dataclass(frozen=True)
class Deliberation:
question: str
positions: dict # model id -> its answer, kept whole
agreed: list # claims every model made
contested: list # claims where the models diverged
# There is no consensus field. Collapsing positions into one answer
# would discard the contested list, which is the part worth reading.Illustrative shape. The absence of a consensus field is the design decision, not an omission.
How this differs from the ordinary version
Disagreement is preserved, not resolved
Ensembling normally means voting, which throws away exactly the signal that was worth having. Keeping the split is what makes the output usable for deciding where to look next.
In the field
Where it earns its cost
Questions with real consequences and no cheap ground truth: a policy reading, an architecture choice, a claim about a regulation. Those are the cases where a single confident answer is the most expensive possible output.
Questions
- Does it pick a winner?
- No. It reports where the models converged and where they did not, and the divergence is the part that is meant to be read.