Series 1 — Part 8 of 8

The shared behavioral ontology is the behavioral AI platform's cross-engine vocabulary. Without it, engines drift into using the same words to mean different things. With it, every engine contributes to a consistent, version-controlled, auditable view of behavioral reality.

What an Ontology Does in a Behavioral System

In a system with 34 engines, the word "trust" appears in at least five contexts: trust graph engine, counterparty trust adapter, negotiation BATNA engine, power dynamics engine, and the governance safe-language map. Without a shared definition, each context evolves independently. Scores become incomparable. Audit trails become meaningless.

The behavioral ontology provides a single canonical definition for every concept, versioned alongside the codebase.

The Ontology Structure

const BehavioralOntology = {
  version: '2.1.0',
  concepts: {
    'trust': {
      definition: 'The assessed likelihood of consistent, predictable behavior aligned with stated commitments.',
      relatedConcepts: ['credibility', 'consistency', 'authority'],
      safeOutputPhrases: ['reliability signals', 'commitment consistency indicators'],
      dimensionType: 'relational',  // 'relational' | 'state' | 'trait' | 'event'
    },
    'behavioral-entropy': {
      definition: 'A measure of unpredictability in an actor\'s decision-making patterns over time.',
      relatedConcepts: ['consistency', 'volatility', 'adaptability'],
      safeOutputPhrases: ['pattern variability', 'decision-making consistency range'],
      dimensionType: 'state',
    },
  },
};

// Engines validate their output labels against the ontology at registration time
function validateEngineLabels(engineDef, ontology) {
  for (const label of engineDef.outputLabels ?? []) {
    if (!ontology.concepts[label]) {
      throw new Error(`Engine ${engineDef.engineId} uses undefined concept: ${label}`);
    }
  }
}

Adaptive Ontology Evolution

The adaptive ontology evolution engine observes engine outputs over time and flags concepts that are being used in divergent ways. When two engines claim to measure "trust" but their outputs are uncorrelated at the 90th percentile, the engine raises a ontology.divergence.detected event — prompting a review.