Definition
Plain language
A series of test runs where you mention an attachment without ever actually attaching one, to see if the mention alone changes the answer.
As stated in the literature
Ablation design that incrementally adds only the linguistic assertion of an attachment (no image tokens present), isolating the effect of the claim from the effect of the modality itself.
Also called: placebo ladders
Why it matters: It separates the effect of an actual image from the effect of merely claiming one exists, which is the only way to know whether a model is responding to the picture or to the words about it.
For example, you send the same question three ways — plain, then "see attached image," then "the attached photo shows a diagram" — while never attaching anything at all.
Heard on the show
“The placebo ladder and the sign inversion are harder to explain away as risk-tracking.”Episode 277 — The Blank White Square That Swings AI Refusal Rates Fifty Points