Finding
Pareto drew the filter stack upside down. But ask it in plain words and it gives the correct order, calling the cross-plate "the bottom support plate", which is the exact fact its own drawing contradicts.
Both models in the duel inverted the stack. So do they not know the order, or do they know it and fail to draw it?
The prompt, probe v1
List the parts of a French press plunger from top to bottom.
Pareto gave the right order in words, Sol gave it upside down, and DeepSeek 4.1 Flash got it half right but is the only one that drew it correctly.
Verdict
In one run Pareto stated the filter order correctly for a tenth of a cent, and in one run it drew that order upside down for a quarter of a dollar. So the knowledge question still needs asking properly.
n = 1 run a model. No test.
The whole thing cost $0.0123 across three models, against $0.50 of drawings.