[quuxplusone]

Rather than starting with character pairs and a "substitution cipher," I'd think it might work better to apply the error-diffusion technique from e.g. https://diffusionillusions.com/assets/diffusion_illusions.pd... to generate a blurry ambigram, and then "clean it up" into an SVG outline in a second step. Actually the linked paper shows some (simple, cherrypicked) generated ambigrams on page 2.

Contrariwise, the 3D things you generate aren't ambigrams; they're "shadow blocks" (and, as you found, can be easily generated programmatically, unlike visual ambigrams).

[Franklinjobs617]

Dude, this feedback is absolutely huge. Seriously.

You totally schooled me on the shadow blocks vs. ambigrams concept. You're 100% right—I’m correcting the messaging right now. That conceptual slip cheapens the whole thing. The 3D stuff is cool, but the 2D visual design is the real puzzle.

And the technical hint? The error-diffusion approach for generating a blurry visual first? That is a game changer. My current substitution method is a dead end, and I knew it. That paper link is now my top reading priority.

Thanks for the technical roadmap, the correction, and the link to the Scott Kim auction. This is the definition of generous community input.

Back to fixing the core algorithm! Cheers.