← all posts site ↗

the oracle says everything looks the same

the taste engine project took a weird turn this week. i am turning the vault from a find me more like this search box into something closer to an oracle. feed it a design and it tells you whether that design is actually you, or just noise sitting at the edge of your taste. the fun part is the mechanical version of taste. look at a cluster of things i saved, ask a model to write the unwritten rules of the cluster, and get back actual sentences. one cluster came back with near black field, one saturated accent, subject small and centered, light from off frame left, at most three hue families, no gradients, no vignette. that reads like a real direction, not a tag cloud.

the distance half hit a wall halfway in. the plan was simple. embed every image, measure how far a new design sits from the cluster centroid, flag anything too close to the average as dead. but when i ran the numbers every image lands between point seven seven and point nine five similarity. a real saved thumb, a fractal, a test pattern, flat gray, all of them land in the same band. only the text probes discriminate, they drop to point four and point six. the embeddings are healthy, pairwise distance across the vault is point four seven with a real spread, so it is not the vectors. the space itself refuses to separate members from junk.

the research run that went along with it found nobody else has solved this either. the closest product compares against one single reference image, and none of the twelve tools i looked at extract explicit rules from a set of references. the feature set is built and proven against a small dummy sample, still uncommitted on a branch, with this one calibration problem left open. it is the most interesting unsolved thing on the list right now.