curiositech/fipa-00086-ontology-service

SKILL: FIPA Ontology Service — Semantic Coordination for Multi-Agent Systems

Decision Points

Mismatch Type → Coordination Strategy

Agent A & B have semantic mismatch:

├─ Syntactic mismatch (different formats/protocols)?
│  └─ NOT an ontology problem → use protocol adapters/formatters
│
├─ Same ontology by reference?
│  └─ SUCCESS → direct communication, no translation needed
│
├─ Different ontologies, assess relationship:
│  ├─ Equivalent conceptualizations, different vocabularies?
│  │  └─ Create bidirectional term mapping → strong translatability
│  │
│  ├─ One ontology subsumes the other?
│  │  ├─ Task requires specialized concepts → use richer ontology
│  │  └─ Task works with general concepts → use simpler ontology
│  │
│  ├─ Partial overlap with acceptable information loss?
│  │  ├─ Task tolerates approximation → implement lossy translation
│  │  └─ Task requires precision → negotiate shared ontology or mediator
│  │
│  └─ No meaningful translation possible?
│     ├─ Find mediating ontology both can translate to
│     ├─ Renegotiate task to use compatible concepts
│     └─ ESCALATE: coordination impossible at ontology level

Explicit vs Implicit Ontology Choice

System requirements analysis:

Will unknown agents join at runtime?
├─ YES → EXPLICIT ontologies required
│  └─ Need discovery/negotiation infrastructure
│
└─ NO → Analyze further:
   │
   ├─ Performance critical (real-time, embedded)?
   │  └─ IMPLICIT → hard-code shared conceptualization
   │
   ├─ Requirements likely to evolve?
   │  └─ EXPLICIT → enable runtime adaptation
   │
   └─ Stable, closed system?
      └─ IMPLICIT → optimize for execution speed

Integration Approach Selection

Integrating existing heterogeneous systems:

Proposed approach is "find ontology intersection"?
├─ YES → RED FLAG: Bottom-up integration
│  ├─ Syntactic overlap ≠ semantic overlap
│  ├─ Risk of silent semantic failures
│  └─ ALTERNATIVE: Top-down foundation approach
│
└─ NO → Assess current approach:
   │
   ├─ Starting with shared conceptual primitives?
   │  └─ GOOD → verify intended models align
   │
   ├─ Mapping to existing standard ontology?
   │  └─ GOOD → validate translation completeness
   │
   └─ Building custom mediating ontology?
      └─ VERIFY: covers both systems' essential concepts

Failure Modes

1. Rubber Stamp Translation

Symptoms: Systems exchange messages successfully, but agents take incompatible actions based on "translated" content Detection Rule: If agents claim successful communication but exhibit contradictory behaviors, check for semantic misalignment despite syntactic compatibility Root Cause: Assumed same vocabulary means same conceptualization Fix: Verify intended models, not just formal models; implement semantic validation checks

2. Intersection Illusion

Symptoms: Integration project finds "common concepts" between ontologies but coordination fails mysteriously during testing Detection Rule: If ontology intersection contains terms with identical names but agents interpret them differently, this is intersection illusion Root Cause: Bottom-up integration assuming syntactic overlap guarantees semantic overlap Fix: Switch to top-down approach; establish shared conceptual foundation first

3. Perfect Translation Trap

Symptoms: Integration project stalls because "perfect" translation between ontologies is impossible to achieve Detection Rule: If project blocks on achieving lossless bidirectional translation when task only requires approximate coordination Root Cause: Over-engineering translation requirements beyond task needs Fix: Match translation quality to task requirements using hierarchy (identical → equivalent → strongly translatable → weakly translatable → approximate)

4. Silent Semantic Drift

Symptoms: System appears to work initially but produces increasingly incorrect results over time Detection Rule: If coordination quality degrades without syntax errors, check for conceptualization divergence Root Cause: Implicit ontologies evolving independently without coordination Fix: Make ontologies explicit; implement semantic version control and change notification

5. Ontology Service Neglect

Symptoms: Each agent implements its own translation logic, creating N² complexity and inconsistent translations Detection Rule: If translation code is duplicated across agents and produces different results for same inputs Root Cause: Treating ontology coordination as per-agent concern rather than system infrastructure Fix: Implement centralized Ontology Agent service; standardize on OKBC interlingua for meta-knowledge operations

Worked Examples

Example 1: Weather Data Integration

Scenario: Emergency response system needs weather data from two sources: MeteoService (European, Celsius, wind in km/h) and WeatherAPI (US, Fahrenheit, wind in mph).

Symptom: Both services provide "temperature" and "wind_speed" but emergency system makes wrong evacuation decisions.

Diagnosis Process:

1. Check syntactic compatibility ✓ (both use JSON)
2. Check ontology relationship: Different vocabularies, same conceptualization (weather measurements)
3. Assess translation requirement: Emergency system needs "hot enough to cause fire risk" (approximate translation sufficient)

Decision Navigation:

• Mismatch type: Different ontologies, partial overlap with acceptable loss
• Task tolerates approximation: YES (fire risk threshold has safety margins)
• Strategy: Implement lossy translation with explicit bounds

Solution: Create translation service with conversion functions and explicit error bounds (±2°C, ±5km/h). Emergency system uses "hot enough for fire risk" threshold that accounts for translation uncertainty.

Expert vs Novice:

• Novice would try to achieve perfect translation or assume same field names mean same semantics
• Expert recognizes approximate translation suffices for task and implements explicit error bounds

Example 2: Multi-Organization Agent Coordination

Scenario: Healthcare agents from Hospital-A (using HL7 FHIR) and Research-Lab (custom medical ontology) need to coordinate patient data for clinical trial.

Symptom: Agents exchange messages but Research-Lab misinterprets "patient status" leading to wrong treatment group assignments.

Diagnosis Process:

1. Both claim FHIR compatibility but Research-Lab extends concepts differently
2. Check conceptualization alignment: Different interpretations of "patient status" (Hospital: administrative, Research: clinical)
3. Translation hierarchy assessment: Weakly translatable (Hospital→Research works, reverse loses information)

Decision Navigation:

• Mismatch type: Extension relationship with semantic drift
• Information loss acceptable: NO (clinical trial requires precision)
• Strategy: Negotiate shared subset or use mediating ontology

Solution: Deploy Ontology Agent with mediating vocabulary covering both systems' essential concepts. Both agents translate to/from mediator rather than direct translation.

Trade-offs Discussed:

• Performance cost of mediated translation vs. risk of semantic errors
• Complexity of three-way ontology maintenance vs. benefits of precision
• Option to constrain Hospital system to Research concepts vs. developing comprehensive mediator

Quality Gates

• [ ] Ontology relationships formally verified (not assumed based on syntax similarity)
• [ ] Translation completeness validated for task-essential concepts
• [ ] Semantic failure detection mechanisms implemented (not just syntactic error handling)
• [ ] Error bounds established for approximate translations with explicit uncertainty propagation
• [ ] Intended model alignment confirmed through test cases, not just formal model compatibility
• [ ] Runtime ontology negotiation protocols specified if system supports dynamic agent joining
• [ ] Escalation paths defined for untranslatable semantic mismatches
• [ ] Performance impact of explicit ontologies measured against coordination benefits
• [ ] Conceptualization vs ontology vs knowledge base layers explicitly distinguished in design
• [ ] Silent semantic failure prevention measures in place (validation, bounds checking, compatibility verification)

NOT-FOR Boundaries

Do NOT use this skill for:

• Syntactic data format conversion → Use data mapping/ETL tools instead
• Database schema integration → Use database integration patterns instead
• API versioning and compatibility → Use API evolution strategies instead
• Performance optimization of knowledge queries → Use database optimization techniques instead
• Natural language processing and understanding → Use NLP/NLU frameworks instead
• Closed systems with stable, shared conceptual models → Use direct integration patterns instead
• Real-time systems where translation overhead is prohibitive → Use pre-compiled shared ontologies instead

When to delegate:

• For syntactic transformations: Use data-transformation-pipelines skill
• For performance-critical knowledge operations: Use knowledge-base-optimization skill
• For natural language semantic understanding: Use semantic-nlp-processing skill
• For API design and evolution: Use api-design-patterns skill

This skill focuses specifically on coordination between autonomous agents with different conceptual models. It is not a general-purpose semantic technology or data integration solution.