On making judgment legible

Discussions of taste collapse three distinct phenomena. Separating them is the first act of intellectual hygiene.

Preference is disposition under choice: coffee over tea, spare over ornate. It need not be justified. It is real, and it is not enough to build institutions on.

Skill is trained perception. The editor who hears a false note in a sentence; the designer who feels density wrong by two percent; the scientist who smells a hand-wavy figure. Skill can be wrong, but it is not arbitrary. It answers to practice, peers, and outcomes.

Signaling is taste as social currency: affiliation, exclusion, status. It rides on the first two and often pretends to be either pure preference (“just my vibe”) or pure skill (“objectively better”). Much of what the internet calls taste is signaling with better fonts.

An API for taste must be honest about which layer it encodes. The interesting target is skillful preference under a declared standard—aspirational where needed, empirical where possible, never confused with universal beauty.

Generative systems produce candidates. Judgment selects among them. The modern stack inverted prestige: generation became the spectacle; judgment became the unpaid intern. That inversion is unstable. When candidates are infinite, selection is the scarce act—and selection without a standard is fashion with latency.

To “have taste” in a productive sense is to possess a standard stable enough to guide work and flexible enough to grow. The standard is partly tacit. Polanyi’s point stands: we know more than we can tell. The engineering response is not to deny tacitness, but to elicit it—through exemplars, counterexamples, comparative rulings, and constitutions written under pressure.

Call the durable part of a taste profile its constitution: a short statement of what the standard seeks and refuses. Constitutions are not slogans. They are constraints on attention—“prefer concrete nouns,” “no fake urgency,” “engineered beauty over decorative futurism.”

Good constitutions are falsifiable in practice. If everything still passes, the constitution is ornamental. If nothing passes, it is pure asceticism or pure snobbery. The laboratory of work—shipping, editing, rejecting—is where constitutions earn their keep.

Versioning matters. Taste drifts; people grow; brands rot and recover. A profile without history cannot be audited. Diff a constitution the way one diffs a legal code: what changed, who changed it, what cases forced the change.

Multiple standards can be serious. That does not make every standard equal. Within a domain—scientific figures, classical performance, typographic craft—communities maintain norms that survive contact with reality. Across domains, standards may be incommensurable. The API does not dissolve that fact; it localizes it. A score is always relative to a profile.

Relativism says: standards are arbitrary, so stop arguing. Pluralism says: standards are situated, so argue carefully—and name the court you are in. Naming the court is what a taste profile does.

To encode taste is to create power: filters on hiring artifacts, brand voice, recommendation, education. Externalization can liberate (a junior designer learns a house standard without telepathy) or ossify (a dead founder’s preferences rule forever). Design accordingly.

Ownership should default to the person or institution whose standard it is. Inference without consent—scraping someone’s corpus into a shadow profile—is a different product and a different ethics. Citation is not optional when profiles inherit from schools, artists, or mentors.

Opacity is a hazard. A score without reasons is a black box with manners. This scaffold insists on breakdowns and explanations even when the underlying judge is crude. Stronger judges must not abandon that obligation.

“High taste” is often a costume: difficulty as virtue, obscurity as proof. That is signaling. The alternative is not populism-as-virtue either. It is seriousness—care about fit between form and purpose; willingness to be wrong in public; loyalty to the work rather than to the club.

An API for taste, if it works, should make seriousness cheaper to maintain and harder to fake. Profiles that only perform erudition will score empty work highly and reveal themselves. Profiles that encode real discrimination will hurt—productively—when the draft is lazy.

We are not proposing that machines become people of taste. We are proposing that people of taste stop accepting tools that cannot hear them. Legibility is not the enemy of mystery; it is the condition of collaboration at scale.

The rest is engineering and practice: better judges, better elicitation, better evals, better manners about ownership. The philosophy only draws the boundary of the problem so the work can begin without self-deception.