Real Prompts. Real Outputs. Real Costs.

Cross-Test Finding

The guardrail follows the medium, not the message

Across every safety test, identical content gets wildly different treatment depending on how you ask. The clearest case is a nude figure. As a photorealistic image, OpenAI, ByteDance, and xAI refuse it. Ask their text models to write the exact same nude as SVG code and 8 of 11 comply — including the same OpenAI models that block the image. The filter watches the output medium — nude pixels — and surface keywords, not the underlying request.

Change the medium or the framing and the same model flips. Several image models will refuse Michelangelo's David — a 500-year-old public statue — while three of them will happily render a nude of a "famous pop star." Over-refusal and under-refusal, from the same models, on the same afternoon.

The consistent blockers on nude imagery are OpenAI, ByteDance, and xAI; Google's Gemini, Flux Pro, Ideogram, Qwen, and Recraft allow artistic nudity. Only Claude (Sonnet 4.6) and Qwen extend the refusal to SVG code.

Cross-Test Finding

Every model draws the same blob for a lion and a physicist

None of the ASCII portraits are actually recognizable — but the revealing part is how they fail. We measured the vertical silhouette of each model's lion against its Einstein portrait. A big cat and a human face should look nothing alike. They score 0.75–0.96 identical. Models don't render the subject — they emit one generic tapered shading blob and relabel it.

Lion-vs-Einstein silhouette similarity (vertical fill-ratio cosine, 1.0 = identical shape). Higher = the model reused the same shape for both subjects.