Ask every way, decide once: making AI decisions invariant to how a question is worded

A decision read off a model's probabilities—a classifier's score, a language model used as a judge, a typed-decision model—depends on how the question is worded. Option order and naming have label-free fixes; logical form has none. We ask each proposition under logically equivalent forms and project the answer vector onto the coherent set, the polytope of vectors consistent with one distribution over the labels. Four label-free facts follow: the decision is invariant over the queried forms; the projection is never worse under the Brier score, whatever the truth; the Brier loss due to form dependence is certified exactly when the projection is interior; with balanced forms the projected decision is the mean-estimate decision. On four public batteries a hosted typed model's single-form rules disagree on 12.5% and 11.8% of items against 0.6% and 1.3% between identical repeats; projection removes the disagreement, lowers the Brier score on every item by exactly its certificate, and leaves accuracy without a statistically clear change. On three open typed models, whose negations are read as affirmations, a label-free additivity weighting of the forms is the best of four poolings on every subsequent run. At equal budget no pooling beats one call.