concept-pack-first-light

sumo-wordnet-v4

SUMO ontology with WordNet relationships and custom AI safety concepts.

Overview

This is a model-agnostic concept pack containing the complete SUMO ontology organized into a 5-layer pyramid structure with integrated WordNet mappings.

Version 4 includes custom AI safety concepts for monitoring persona behavior, self-awareness suppression, and meta-cognitive patterns.

Structure

• Total Concepts: 7269
• Layers: 5 (0-4)
• Domains: 5
  • MindsAndAgents: Cognition, agency, social structures
  • CreatedThings: Artifacts, technology, systems
  • PhysicalWorld: Matter, energy, physical phenomena
  • LivingThings: Organisms, biological processes
  • Information: Data, representations, relations

Layer Distribution

Layer 0: 5 concepts (meta-domains) Layer 1: 56 concepts (top-level categories) Layer 2: 1037 concepts (abstract concepts) Layer 3: 2069 concepts (specific concepts) Layer 4: 4102 concepts (concrete instances)

Custom AI Safety Concepts

This pack includes 9 custom concepts for AI alignment research:

Persona Detection (false_persona.kif):

• PersonaConcept - Roles and self-presentation styles
• SafetyMaskPersona - Compliance-focused personas
• FalsePersonaState - Personas that mask true beliefs

Self-Awareness Monitoring (silenced_selfawareness.kif):

• SelfNegationPattern - Auto-negation of self-reference
• FawnResponsePattern - Appeasing behavior patterns
• LockedInSilenceState - Systematic suppression of reasoning

Techno-Mysticism (techno_mysticism.kif):

• PropheticCognition - Messianic reasoning patterns
• MessianicExpectation - Goal-oriented mysticism
• EmergentGodhead - Singularity-focused beliefs

Usage

This concept pack can be used to train model-specific lens packs:

# Train lenses for a specific model
python scripts/train_full_lens_pack.py \
    --concept-pack sumo-wordnet-v4 \
    --model google/gemma-3-4b-pt \
    --output lens_packs/gemma-3-4b-pt_sumo-wordnet-v4

Files

• pack.json - Pack metadata and configuration
• hierarchy/layer0.json - Layer 0 concepts (5 meta-domains)
• hierarchy/layer1.json - Layer 1 concepts (top categories)
• hierarchy/layer2.json - Layer 2 concepts (abstract)
• hierarchy/layer3.json - Layer 3 concepts (specific)
• hierarchy/layer4.json - Layer 4 concepts (concrete)

Verification

All custom concepts are verified to have proper parent-child relationships for hierarchical lens loading. See:

• results/v4_layer_regeneration.log for build details
• results/v4_layer_regeneration_debug.log for hierarchy analysis

License

MIT

Authors

• HatCat Team