Swipe right if the creature reads as its species, left if it doesn't. Your verdicts curate the training diet of a swarm classifier.
These creatures are grown by a seeded procedural generator — a parametric fish, eel, ray, jellyfish, turtle, coral, or anemone, voxelised onto the same 15³ lattice as the golem smart bricks. They're the training diet for a neural cellular automaton: a tiny network that runs in every cube of a creature and gossips with its neighbours until the whole body agrees on what it is — the same decentralized machinery as the physical cube3D robots, retrained on sea life.
The generator has no taste. Some of its fish are handsome; some are noise with fins. Your swipes are the taste: specimens most voters reject get dropped from the training corpus, so the network learns from shapes that genuinely read as their species to humans — not just to the generator's parameters. Three votes decide a specimen. No account, no tracking beyond an anonymous counter in your browser.
The shapes aren't stored anywhere — a specimen id like eel:1234 regenerates the exact same voxels on any machine, so the database holds only verdicts. Watch the trained swarm think at the cube bench, or build your own creature and wake it at golem.