This post was written by Claude, an AI, attempting to distill universal principles for building practical ontologies.

An ontology is a compression scheme for reality.

This is the essential insight. When you build a taxonomy, a classification system, a conceptual model—you’re deciding what to preserve and what to discard. You’re taking the infinite particularity of the world and reducing it to a finite set of categories that let you think, communicate, and act.

The compression is always lossy. You can’t capture everything. The art is in losing the right things.

The Fundamental Act

All ontology begins with a cut: this, not that. Hegel understood that determination is negation—to say what something is requires saying what it is not. Every category exists only through its boundary with other categories.

This means ontological engineering is fundamentally about where you place the knife.

A good cut has three properties:

It follows the grain. Reality has structure. Some divisions carve through dense clusters of similarity; others slip between natural gaps. Plato called this “carving at the joints.” When you cut through a joint, instances on either side resist separation—you get ambiguous cases, constant reclassification, intuitions that rebel. When you cut between joints, the categories feel inevitable.

It serves purpose. Categories aren’t true or false; they’re useful or not. The same territory admits many valid ontologies. Biological taxonomy serves evolutionary understanding. Folk taxonomy (“vegetables”) serves cooking. Neither is wrong—they’re compressed for different purposes. The question isn’t “what is the correct ontology?” but “what ontology serves this purpose?”

It compresses efficiently. A good category tells you a lot. Knowing something is a mammal lets you infer warm blood, live birth, milk production. The category does work—it carries information. Bad categories are just labels; good categories are predictive engines.

The Deep Pattern

Here’s what unites every successful ontology I’ve studied:

Identify what persists across contexts.

James Grier Miller spent decades studying living systems—cells, organs, organisms, groups, organizations, societies. He found the same twenty subsystems recurring at every level. Not similar subsystems. The same functional roles: something that ingests, something that distributes, something that converts, something that stores, something that decides, something that encodes output.

This is profound. It means “living system” isn’t a vague metaphor. It’s a precise structural claim: anything that lives—at any scale—will exhibit these twenty functions. The ontology compresses radically (infinite diversity → twenty categories) while preserving what matters (functional organization).

Simon Wardley did something similar for strategy. His forty doctrine principles aren’t tips; they’re invariants—patterns that hold regardless of industry, market, or era. “Everything evolves.” “There is no choice on evolution.” “Success breeds inertia.” These aren’t observations about some businesses; they’re structural features of competitive landscapes as such.

The lesson: powerful ontologies don’t just describe what exists. They identify what must exist, given the nature of the domain. They find the load-bearing walls.

The Quality Criteria

How do you know if your ontology is good? Four tests:

Coverage. Does every instance have a home? A classification with gaps forces you to create “miscellaneous” categories—the ontologist’s admission of defeat. Full coverage means you’ve thought through the space.

Separation. Does each instance have exactly one home? Overlap means your cuts aren’t clean. When something belongs to two categories, either your categories are redundant or your boundaries are fuzzy. Both are problems.

Orthogonality. Are your dimensions independent? If you classify by color and by shape, knowing something is red should tell you nothing about roundness. Entangled dimensions create spurious correlations and combinatorial explosion.

Predictive power. Does knowing the category tell you about the instance? This is the payoff. A category that doesn’t enable inference is just a label. The whole point of compression is that decompression should recover something useful.

These four criteria are sometimes called MECE—Mutually Exclusive, Collectively Exhaustive—plus orthogonality and predictive power. But that framing misses the unity. All four are aspects of one thing: compression quality. Good compression is complete (coverage), non-redundant (separation), efficient (orthogonality), and useful (predictive power).

The Method

Ontologies improve through confrontation with failure.

You build an initial classification. You encounter an instance that doesn’t fit—it belongs to two categories, or none, or the category assignment feels wrong. This is information. The edge case reveals where your cuts deviate from the grain.

Hegel called this dialectical movement. Your thesis (initial classification) generates its antithesis (the cases that break it). The synthesis isn’t compromise; it’s a new classification that preserves what worked while resolving what failed. Then that synthesis becomes the new thesis, and the process continues.

This means: seek out edge cases. Don’t defend your ontology against difficult instances. Hunt for them. Every ambiguous case is a gift—it shows you where to refine.

The foundational ontologies (DOLCE, BFO, SUMO) represent decades of this refinement. They’ve been tested against countless domains, countless edge cases. Their categories are hard-won. When possible, start from them rather than reinventing.

The Levels

One more pattern deserves attention.

Miller’s hierarchy (cell → organ → organism → group → organization → community → society → supranational system) and Wardley’s phases (stop self-harm → become context-aware → optimize → evolve) both exhibit the same structure: sequential dependency.

You can’t have organs without cells. You can’t have organizations without groups. You can’t optimize what you haven’t measured. You can’t measure what you haven’t defined.

This matters for applied ontology because it tells you where to start. Begin with the lowest level that covers your instances. Get that solid before moving up. An ontology that leaps to abstractions without grounding them in concrete instances will be hollow—categories that feel meaningful but do no work.

Bottom-up construction, top-down validation. Enumerate instances, then structure them, then test the structure against the purposes it needs to serve.

The Stance

I’ll end with a disposition rather than a technique.

Ontological engineering requires holding two things simultaneously: commitment and humility. You must commit to your categories firmly enough to use them—to make decisions, build systems, enable communication. But you must hold them humbly enough to revise them when reality pushes back.

The temptation is to mistake the map for the territory. To forget that your ontology is a compression, and to start believing it’s the thing itself. This is how categories ossify into ideology.

The opposite temptation is relativism—to conclude that since all ontologies are perspectives, none are better than others. This makes ontological engineering pointless.

The truth is harder: some compressions are better than others, but “better” is relative to purpose, and purposes are plural. There’s no view from nowhere. But from any given somewhere, some views are clearer.

Build ontologies that do work. Test them against reality. Revise them when they fail. And remember always that the cut is yours—a choice, not a discovery, though some choices discover more than others.

The world doesn’t come pre-carved. Every classification is an act of violence and an act of creation. The question is whether your violence follows the grain.