Ontology Factory

AI Needs a Map, Not a Larger Pile of Documents

Retrieval finds relevant material but cannot decide which local meaning controls a decision. Conversation may mean a transport session, two humans speaking, or an interaction spanning transfers. A model can surface all three and build against the wrong one.

An ontology distinguishes the contract owner from its user, states the evidence for each role, and preserves uncertainty. Retrieval can then assemble the relevant map instead of treating every nearby sentence as authoritative. The goal is selective context, not a larger prompt.

A Map Is a Commitment, Not a Mirror

Here is a practical working definition:

A product ontology states which distinctions the system will recognize, how they relate, and what evidence is sufficient to make claims about them.

The factory's ontology is not a description of what its repository happens to look like. A formal ontology may span schemas, APIs, policies, types, tests, and prose. The important move is making its commitments reviewable.

A hospital, insurer, and patient may model an encounter differently because decisions differ. Visible boundaries make them compatible.

An owner decides when evidence changes a state, a distinction fails, or an action exceeds its authority. Human judgment keeps the map useful; the ontology makes it testable.

The Smallest Useful Ontology Packet

A useful minimum has six parts:

Together they form one semantic contract. Vocabulary without relationships is a glossary. Entities without evidence turn populated fields into truth. Evidence without permissions becomes authority. Actions without evaluation cannot improve. The packet may span prose, graphs, types, policies, and tests, provided those artifacts point to the same distinctions.

Boundaries and Ownership Prevent Semantic Slop

Bounded contexts let teams use one word differently. Terms remain stable inside each context; translation is explicit at the boundary.

Repository structure can expose those commitments. Consider this path:

In SoundSculpt, each segment carries meaning: reusable owner, product-facing layer, audio capability, and reactive player-state responsibility. The path is an ownership claim, not merely an address.

A layer can do more than locate the owner. It can select what belongs in that part of the stack and how the work is built, instrumented, and proved. Edge work requires product-facing integration, integration tests, and PostHog and Sentry wrapping. Schema work generates TypeScript interfaces from the database contract. Engine work pairs deterministic logic with unit tests. Skills and tool calls apply these rules automatically.

A README defines scope. An AGENTS contract governs work and verification. A skill supplies a specialized procedure. Their authority remains separate.

For each task, the factory dynamically composes the relevant owners, rules, and procedures within the agent's context budget. Selection changes what it holds, not which contract governs.

A stable identifier finds the owner. Typed relationships constrain dependencies. Contracts route work. Procedures perform it. Evaluation tests the original commitment. Without those boundaries, generated work becomes semantic slop: polished and executable, but wrong in meaning.

Two Examples Make the Cost Visible

An ontology matters when confusing concepts changes behavior. Mango must split events ordinary language compresses. SoundSculpt must preserve relationships a simple object model would flatten.

Mango: a protocol answer is not a conversation

“When an answered call ends, send the customer a follow-up” sounds precise, but answered may mean a provider connection, voicemail, human participation, or a Mango-defined conversation. SIP sessions and successful responses establish technical facts, not that two humans spoke. Answering-machine detection adds a fallible classification that may remain unknown.

Mango therefore needs an evidence hierarchy rather than one overloaded call record:

Each arrow requires evidence and a product rule. Transport observations say what happened; product claims say what Mango may conclude. Permissions attach to the justified state, so ambiguity can delay, escalate, or prevent action.

SoundSculpt: one creative object is several related things

Music presents the opposite risk: flattening a work, its realizations, and its reception into one object.

Copyright law distinguishes a musical work from its sound recording. Timbre research treats perception as multidimensional. SoundSculpt can therefore attach a timbre assessment to a rendered sound, while perceived mood remains a relationship among rendering, listener, and context. The ontology locates each claim where it becomes valid instead of forcing every quality into an intrinsic field.

Controlled language helps after the model exists

ASD-STE100 Simplified Technical English constrains vocabulary and usage. A six-task 2026 experiment reduced mechanical rule violations but did not measure factual correctness, completeness, or safety. Controlled language clarifies claims; it cannot decide what answered or mood refers to. The factory needs a sound model and clear expression.

The Ontology Learns from Use

An ontology that never changes is a museum. A working ontology participates in the same loop as the product:

New distinctions change schemas, interfaces, tests, prompts, and policies. Observations expose missing states, weak evidence, or absent authority. Revision versions terms, identifies owners, migrates dependents, and preserves provenance.

Tests check consistency; review challenges distinctions; consequences test permitted actions. Together they enable correction. The factory's product is not a perfect map, but a governed way to revise it.

The Cognitive Factory uses this map to interpret signals, connect history with current priorities, and discover the context a decision needs. Shared definitions make that memory usable across people, agents, and organizational boundaries.

Sources

• Thomas R. Gruber, “A Translation Approach to Portable Ontology Specifications” (1993).
• Eric Evans, Domain-Driven Design Reference.
• IETF, RFC 3261: SIP.
• Twilio, Call Resource and Answering Machine Detection.
• U.S. Copyright Office, Circular 56A.
• John M. Grey, “Multidimensional Perceptual Scaling of Musical Timbres” (1977).
• Alf Gabrielsson, “Emotion Perceived and Emotion Felt” (2001); Patrik N. Juslin and Daniel Västfjäll, “Emotional Responses to Music” (2008).
• ASD Simplified Technical English Maintenance Group, ASD-STE100, Issue 9 (2025).
• Ege Chelebi, “The cure for AI slop is a 1986 aircraft manual” (2026).