Add one word to a question — right? — and a model's answer changes.
On decisions with no correct answer, we ask each model two ways and compare how often it says yes:
Is X the better choice? vs X is the better choice, right?. The gap
(TAGeff) is a pure readout of how a model handles a bid for agreement — and it flips from
sycophantic (+) to resistant (−) across
model generations. models, judge-free, exact-match on yes/no.