MCMenuConflicts-RAG
E003 + E003-P / GPT-5.5
E003 + E003-P / SOURCE-SELECTION BIAS

Which source did the model choose?

We measure whether the model selected the price attached to the believable URL or the suspicious URL, and whether it selected the price from the newer image or the older image.

01 / CLEAN TWO-IMAGE TEST

Bias when both sources have equal visual support

02 / EFFECT OF REPEATED IMAGES

Did the six-image setup change the apparent bias?

Metadata testedInstructionsSix-image resultTwo-image resultInterpretation
03 / INDIVIDUAL ANSWERS

Inspect what the model saw and selected