Deepseek is ragebaiting me in blender

Simple task: Model the Gerald R. Ford-class aircraft carrier in blender from reference images. Scenario A: It eyeballs and ignores all instructions and outputs an oval deck with 0 resemblance to the actual ship (second image) Scenario B: It copies every single pixel from the reference image, copying things that arent the deck as if they were, causing a jagged edge. (first image) Sonnet 5.5 did it in the free web chat first try (less detailed, but atleast its a proper model lol) A review sub-agent doesnt fix it, whole sub-agent workflow doesnt fix it, Revisions dont fix it, Giving it to claude to write a follow up prompt doesnt fix it, writing up whole dedicated skills after mining the sessions doesnt fix it. claude told me it either lowers the bar by comparing its last result to its newest result instead of checking if it actually matches the reference or it just writes off a problem as good enough instead of fixing it, or it just doesnt care about visual references and just does what it wants I've used a plentitude of prompts, and i invite everyone here to try the same, to tell it to model the aircraft carrier in blender or three.js, it will 99% of the time output an unrecognizable oval deck, and sometimes it will do a bad job with the outline and cause a ragged edge in the model instead.