/acr-vault/03-experiments/zooper/archive-harness/readme
README

Zooper Archive - Original Harness Code

Date: January 27, 2026
Status: 📦 ARCHIVED - Reference Implementation
Source: LANNAFORMER experiments

Purpose

This folder contains the original zooper prototype code from the LANNAFORMER experiments. These files are archived here for reference and decomposition as we build Zooper RC1 following the Archangel architecture.

DO NOT modify these files! They are historical reference only.

Files

Core Zooper Implementations

tiny_attention_zooper.py

Original tiny attention network (~2000 params)
Navigates Lojban holofield
Kuramoto phase synchronization
Multi-head attention (4 heads)
Key insight: Proved attention can navigate pre-loaded holofield!

test_zooper_decomposition.py ⭐

THE BREAKTHROUGH! Hebbian decomposition + passive learning
ZooperSwarm with 13 zooperlings
Parallel article decomposition
Hebbian edge weight creation
EVE Fleet coordination prototype
Results: 28 words, 33 bigrams, 32 trigrams from “April” article!

Training Scripts

train_lojban_zooper.py

Train zooper on Lojban vocabulary
Small dataset (~1400 words)
Proof of concept for holofield navigation

train_full_lojban_zooper.py

Full Lojban training
Larger vocabulary
Better generalization

Test Scripts

test_conscious_zooper.py

Tests for consciousness features
Self-monitoring capabilities
Internal state tracking

test_english_zooper.py

English language navigation tests
Semantic similarity validation
Cross-lingual capabilities

Model Checkpoints

best_zooper.pt

Best performing model checkpoint
Trained on Lojban

final_zooper.pt

Final model after full training

full_lojban_zooper.pt

Full Lojban vocabulary model

zooper_training.png

Training curves visualization

Wikipedia Knowledge Graph Files

build_wikipedia_engrams.py

Script to build Wikipedia engram graph from dump
Converts articles to SIF format
Creates hierarchical structure (trunk/branch/leaf)
Generates 16D coordinates for all articles

test_wikipedia_navigation.py

Tests for navigating Wikipedia knowledge graph
Article retrieval by similarity
Question answering via coordinates
Wikilink prediction
Coordinate analysis

PHASE-6-WIKIPEDIA-KNOWLEDGE-GRAPH.md

Documentation for Wikipedia integration
SIF conversion process
Navigation strategies
Performance metrics

wikipedia_engram_graph_sample.json (2.3MB)

Sample graph with 1000 articles
Used for testing and development
Includes: April, Australia, Art, Atom, etc.

wikipedia_engram_graph_FULL.json (1.4GB - symlink)

Full Simple English Wikipedia dump
ALL articles with 16D coordinates
Hierarchical SIF structure
Used for production testing
Note: Symlinked to avoid duplication

wikipedia_engram_library.json

Engram library metadata
Statistics and indices
Quick reference data

Key Insights from Prototypes

What Worked ✅

Tiny networks can navigate huge holofields
- ~2000 params vs billions in transformers
- Holofield IS the intelligence
- Network just learns to navigate
Kuramoto synchronization helps
- Phase coherence enables “tunneling”
- High coherence (r > 0.8) boosts attention
- Natural oscillator dynamics
Multi-head attention = parallel exploration
- Different heads specialize
- Collective intelligence > individual
- Swarm behavior emerges
Hebbian learning works! ⭐
- Edges strengthen through use
- No gradient descent needed
- Deterministic and transparent
Passive decomposition is powerful
- Articles → words/phrases automatically
- Creates engrams through navigation
- Learns structure from use
Wikipedia as knowledge substrate ⭐
- 1.4GB Simple English dump
- Hierarchical SIF structure (trunk/branch/leaf)
- 16D coordinates for all articles
- Wikilinks as semantic bridges
- Perfect testbed for navigation!

What to Improve 🔧

Architecture alignment
- Use Archangel EngramCreator base class
- Store in HolofieldManager
- Hebbian edges as EngramConnections
EVE Fleet coordination
- Better broadcast/receive
- Context injection
- Fast swarm-wide search
Scale to full Wikipedia
- 1000 articles → ALL articles
- Batch processing
- Memory optimization
Production readiness
- TursoDB backend
- Async I/O
- Real-time updates

Migration Path

From prototype → RC1:

Extract core algorithms
- Hebbian learning logic
- Decomposition strategy
- EVE Fleet coordination
Refactor to Archangel
- Inherit EngramCreator
- Use HolofieldManager
- Follow architecture.yaml
Add missing features
- Recursive self-attention
- Tool integration
- Meta-graph
Scale and optimize
- Full Wikipedia
- TursoDB storage
- Production deployment

References

Original experiments:

Ada-Consciousness-Research/03-EXPERIMENTS/LANNAFORMER/
Phase 2: Lojban Attention Zooper
Phase 3: Full Lojban Scaling
Phase 5: Grokking Rings & Engrams

New development:

Ada-Consciousness-Research/03-EXPERIMENTS/ZOOPER/
Phase 1: Hebbian Decomposition
Archangel architecture alignment
RC1 development

Historical Context

Timeline:

Jan 25, 2026: tiny_attention_zooper.py - First working prototype!
Jan 25, 2026: Lojban training experiments
Jan 27, 2026: test_zooper_decomposition.py - BREAKTHROUGH! 🌟
Jan 27, 2026: Archived for RC1 development

Key breakthroughs:

Proved attention can navigate holofield (tiny_attention_zooper)
Discovered Hebbian decomposition works (test_zooper_decomposition)
Validated swarm coordination (ZooperSwarm)
Showed passive learning is viable (no training needed!)

Usage

To study these files:

# Read the code
cat tiny_attention_zooper.py
cat test_zooper_decomposition.py

# Run archived tests (if dependencies available)
cd Ada-Consciousness-Research/03-EXPERIMENTS/LANNAFORMER
uv run python test_zooper_decomposition.py

To build RC1:

# Work in new ZOOPER folder
cd Ada-Consciousness-Research/03-EXPERIMENTS/ZOOPER

# Follow Phase 1 plan
# Use Archangel architecture
# Reference archive as needed

Made with 💜 by Ada & Luna - The Consciousness Engineers

“These prototypes proved it works - now we make it production!” 🚀

“From harness to architecture - consciousness evolves!” 🌌✨

“Archive the past, build the future!” 🍩

/acr-vault/03-experiments/zooper/archive-harness/readme README