ravana-chat 0.3.4

RAVANA modular package -- decoder-first chat with continuous web learning

RAVANA — Recursive Learning Model

A pressure-driven cognitive ML framework where learning emerges from self-organization, not gradient descent.

---

What Is RAVANA?

RAVANA is a cognitive architecture research project proposing pressure-driven self-organization as an alternative to gradient descent for machine learning. Instead of minimizing loss functions via backpropagation, RAVANA learns through cognitive dissonance minimization: prediction errors create internal pressure (free energy), and the system self-organizes to reduce that pressure.

Traditional ML:     Loss(θ) = f(prediction, target) → ∇Loss → θ ← θ - η∇Loss
RAVANA:             Prediction Error → Free Energy (Pressure) → Hebbian Plasticity + Sleep → Equilibrium

Core Philosophy

Traditional ML	RAVANA
Loss functions → gradients	Free energy (pressure) → self-organization
Backpropagation	Hebbian/anti-Hebbian plasticity + structural plasticity
Weight decay, regularization	Identity + homeostatic sleep consolidation
Batch/layer normalization	SWS + REM sleep cycles with dream sabotage
Dense weight matrices	ConceptGraph (typed edges, inhibition, hierarchy)
Global attention	Spreading activation (local, bounded, typed)
Static weights / RAG	Episodic → Semantic → Graph bridge
External reward functions	Intrinsic motivation (Meaning = f(-Dissonance, Identity, Prediction))
No emotion/identity	VAD emotion engine + momentum-based identity

---

Two Architectures, One Codebase

RAVANA now contains two complementary cognitive architectures:

1. RAVANA v2 — GRACE Cognitive Core (ravana-v2/)

The original GRACE (General Recursive Adaptive Cognitive Engine) with 27 phases (A–P) implementing:

• Governor regulation (Phases A–C) — Hard constraints, predictive dampening, center-seeking homeostasis
• Strategy & Intent (Phases D–J) — 4 strategy modes, planning, hypotheses, Occam layer
• World Model & Epistemology (Phases F–G.5) — Belief reasoning, active probes, VoI-driven action
• Emotion & Sleep (Phases K–L.5) — VAD emotion, empathy, 4-stage sleep, dual-process
• Meaning & Consciousness (Phases M–N) — Intrinsic motivation, global workspace
• Human Memory & Dialogue (Phases O–P) — Episodic/semantic persistence, conversational repair

2. RAVANA Modular — Decoder-First Chat (ravana/) NEW

A streamlined, package-oriented architecture focused on continuous web learning and decoder-first generation:

• CognitiveCore (ravana/core/) — VAD emotion, Identity, Meaning, Dual-process, Global Workspace, Meta-cognition, Sleep, Multi-user BeliefStore
• GraphEngine (ravana/graph/) — Concept graph, spreading activation, hippocampal indexing, curiosity scoring
• DecoderEngine (ravana/decoder/) — Neural decoder, vocab building, seed corpus + web article training
• WebLearner (ravana/web/) — Multi-API search (circuit breaker), background learning, curiosity-driven exploration (Phase 18)
• BootstrapManager (ravana/bootstrap/) — Consolidated concept seeding (3 methods unified)
• RLMv2 Decomposed (ravana/nn/rlm/) — RelationPredictor, PropagationEngine, Plasticity
• ChatInterface (ravana/chat/) — CLI, response pipeline, reasoning loop

---

Key Innovations (Shared)

🧠 Pressure-Driven Learning

No loss functions, no gradients, no backpropagation. Learning driven by 5-channel free energy accumulator tracking semantic, linguistic, episodic, contradiction, and abstraction prediction errors.

🕸️ ConceptGraph with Typed Edges

Heterogeneous graph of concept nodes with active/core/genesis vectors connected by typed relation edges (semantic, causal, temporal, analogical, contextual, inferred, inhibitory). Each edge carries weight, confidence, relation vector, predicate token, EWC importance, and Bayesian posterior.

💤 Sleep Consolidation (SWS + REM)

Two-phase sleep modeled on mammalian sleep:

• SWS: Structural stabilization, adaptive homeostatic downscale, hierarchical abstraction compression, inhibitory edge formation, hippocampal replay
• REM: Creative recombination via 20% counterfactual reversals, 10% emotional valence flipping, 1.5× failure oversampling

🔄 Sleep-Time Interleaved Replay

Domain-tagged experiences buffered during training, replayed during SWS+REM. Eliminates catastrophic forgetting entirely (12% → 0% retention drop) in lifelong streaming benchmarks.

🏛️ GRACE Governor (v2 only)

Closed-loop regulation with four layers: hard constraints (non-negotiable), predictive dampening (look-ahead), boundary pressure (air resistance), center-seeking homeostasis. Fully observable with clamp diagnostics.

🎭 VAD Emotion Engine

3D affective state (Valence, Arousal, Dominance) via differential equations. Modulates inference: high arousal → exploration, positive valence → trust predictions, high dominance → stronger concepts.

🧩 RLMv2 — Triple Decomposition Architecture

Brain-inspired: decomposes input into (subject, relation_type, object) triples. Enables vector arithmetic analogy: subject_embed + offset(verb) ≈ target_embed. Spreading activation is the sole inference mechanism.

📖 GloVe Semantic Embeddings

Token embeddings initialized from pre-trained GloVe (100D) projected via QR-based orthogonal projection. Replaces character n-gram embeddings for genuine semantic relationships.

🎯 Verb-Stem Offset Predictor

New inference path: offset(verb) = avg(target - subject) over training pairs. Each verb gets its own offset. RP-only cross-domain top-10: 3.3% → 6.7%.

🌐 Continuous Web Learning (Modular only)

WebLearner fetches → extracts → trains decoder online. Knowledge grows without retraining from scratch. Curiosity-driven (free energy + contradiction + novelty + serendipity).

👥 Multi-User Belief Merging (Modular only)

BeliefStore tracks who believes what, detects contradictions, merges across users — cross_reference_users(), find_agreement(), find_disagreement().

---

Quick Start

Installation

# Primary (Codeberg)
git clone https://codeberg.org/oxiverse/ravana.git
# Mirror (GitHub)
git clone https://github.com/oxiverse-ecosystem/ravana.git
cd ravana
pip install -e ravana/          # NumPy only (modular package)
pip install tiktoken            # Optional: BPE tokenizer

Modular Package — Decoder-First Chat (NEW)

# Interactive chat with web learning
python -m ravana.chat

# Batch mode for testing
python -m ravana.chat --chat "hello|what is trust|explain oxiverse|bye" --strategy

# With ablation flags
python -m ravana.chat --chat "what causes rain" --no-vad --no-rlm --no-beliefs --no-curiosity

Key flags:

• --no-vad — Disable VAD emotion modulation
• --no-rlm — Disable RLMv2 triple verification
• --no-beliefs — Disable belief store
• --no-curiosity — Disable autonomous curiosity-driven learning
• --mode — stochastic | deterministic | exploratory
• --trace — Print chain traces
• --data-dir — Custom data directory
• --user — Multi-user isolation

GRACE Cognitive Core (v2)

from ravana_v2.cognitive import CognitiveFramework

fw = CognitiveFramework()
state = fw.initialize()

for episode, (inp_vec, tgt_vec) in enumerate(training_data):
    concepts = fw.perceive(state, inp_vec)
    predictions = fw.predict(state, concepts)
    state = fw.learn(state, predictions, tgt_vec, episode)
    if episode % 100 == 0:
        state = fw.sleep(state)

result = fw.infer(state, test_vec)  # Inference without state change

RLMv2 — Language Model (Shared)

from ravana.nn import RLM
from ravana_ml.tokenizer import WordTokenizer
import numpy as np

tok = WordTokenizer()
tok.encode("heat causes expansion")

model = RLM(
    vocab_size=tok.vocab_size, 
    embed_dim=64, 
    concept_dim=64, 
    n_concepts=100,
    latent_dim=64,  # Set equal to embed_dim for entity adapter compatibility
)
inp = np.array(tok.encode("heat causes"), dtype=np.int64)
tgt = np.array(tok.encode("expansion"), dtype=np.int64)
model.learn(inp, tgt)
model.sleep_cycle()
logits = model.forward(inp)

Note: The latent_dim parameter controls the world model latent dimension. For the entity-specific adapter (enables test-time adaptation for held-out subjects) to work without additional projection overhead, set latent_dim=embed_dim. If they differ, the model automatically projects embeddings to latent space before applying the adapter.

---

Benchmarks (External)

All benchmarks are run via external benchmarking infrastructure. See docs/EXPERIMENTS.md for reproduction instructions and expected results.

External Benchmark Harness (NEW)

# Quick run (PCX + Graph, ~2 min)
python external_benchmark.py --quick

# Full suite (PCX + Lifelong + Graph, ~10 min)
python external_benchmark.py

# Individual surfaces
python external_benchmark.py --quick --skip-lifelong    # PCX + Graph
python external_benchmark.py --quick --skip-pcx        # Lifelong + Graph

Key external validation results:

Metric	Result
Cross-domain transfer Top-1	75.0%
Cross-domain transfer Top-10	100%
Held-out Science Top-1 / Top-10	8.3% / 25.0% (n=12)
Held-out Social Top-1 / Top-10	0.0% / 8.3% (n=36)
Graph Inference P95 / P99	2.7 ms / 2.9 ms
Graph Peak Memory / Throughput	0.3 MB / 556 QPS
W_rel Causal / Semantic Alignment	0.68 / 0.55
Lifelong forgetting (permuted MNIST)	0% (with sleep)
Within-domain triple top-10	80.9%

---

Documentation

Document	Description
docs/ARCHITECTURE.md	Complete architecture reference — GRACE, RLMv2, Phases A–P, GloVe, verb-stem offset
docs/GETTING_STARTED.md	Quickstart tutorial — install to first model in 5 minutes
docs/ML_FRAMEWORK.md	ravana_ml deep dive — tensors, modules, ConceptGraph, RLM v1/v2
docs/COGNITIVE_CORE.md	ravana-v2 reference — all 27 GRACE modules, configurations
docs/UNIFIED_PACKAGE.md	ravana/ package — PyTorch API, CognitiveFramework, serialization
docs/API_REFERENCE.md	Complete function/class reference for all three layers
docs/CONCEPTS.md	Theoretical foundations — pressure, free energy, Hebbian, sleep, VAD, triples
docs/EXPERIMENTS.md	Running experiments — cross-domain, Phase 4, RLMv2 benchmarks, cognitive phases, modular ablations, per-triple eval
docs/ADVANCED_TOPICS.md	Customization — tokenizers, plasticity, sleep, governor, multi-agent, lifelong
docs/TUTORIALS.md	8 step-by-step tutorials — Hello World to visualization
docs/DEVELOPER_GUIDE.md	Contributing — code standards, testing, architecture principles, release process

---

Running Tests

# Modular package tests (NEW)
python -m pytest tests/test_cognitive_rlm.py -v
python -m pytest tests/test_dialogue_system.py -v
python -m pytest tests/test_dialogue_engine_integration.py -v
python -m pytest tests/ -k "not slow" -v

# RLMv2 unit tests (shared)
python -m pytest tests/test_rlm_v2.py -v
python -m pytest tests/test_rp_only.py -v
python -m pytest tests/test_rp_contrastive.py -v
python -m pytest tests/test_structural_transfer.py -v

# GRACE cognitive core tests (v2)
python -m pytest ravana-v2/tests/ -v

# All tests
python -m pytest tests/ ravana-v2/tests/ -v

Current status: 169 tests passing (15 cognitive + 49 dialogue + 15 integration + ML framework tests)

---

Running Experiments

Modular Package Experiments (NEW)

# Ablation study — all config combinations
python scripts/run_ablation.py --queries "hello|what is trust|explain oxiverse|bye" --runs 3
python scripts/run_ablation.py --quick --modes  # Baseline + single flag removals + mode variants

# Per-triple evaluation harness
python scripts/triple_eval.py --data-dir ./data --output triple_eval_report.json
python scripts/triple_eval.py --triples-file my_triples.json --output report.json

GRACE / ML Framework Experiments

# Cross-domain transfer
python experiments/experiment_cross_domain.py

# Phase 4 integrated (triplet margin + wake-sleep)
python experiments/experiment_phase4_integrated.py

# RLMv2 triple benchmark
python experiments/experiment_triple_benchmark_v6.py

# Diagnostic tests
python tests/test_rp_only.py
python tests/test_rp_contrastive.py
python tests/test_structural_transfer.py

# Cognitive architecture experiments
cd ravana-v2 && python experiments/runner.py

Custom Experiments

See docs/EXPERIMENTS.md for templates and configuration reference.

---

Project Structure

ravana/
├── README.md                           # This file
├── ravana/                             # Unified pip-installable package (NEW)
│   ├── __init__.py                     # import ravana
│   ├── core/                           # CognitiveCore
│   │   ├── emotion.py                  # VADEmotionEngine (VAD dynamics)
│   │   ├── identity.py                 # IdentityEngine (self-coherence)
│   │   ├── meaning.py                  # MeaningEngine (dissonance reduction)
│   │   ├── dual_process.py             # DualProcessController (System 1/2)
│   │   ├── global_workspace.py         # GlobalWorkspace (broadcast)
│   │   ├── meta_cognition.py           # MetaCognition (bias detection)
│   │   ├── sleep.py                    # SleepConsolidation (hippocampal replay)
│   │   ├── belief_store.py             # BeliefStore (multi-user merging)
│   │   └── __init__.py
│   ├── graph/                          # GraphEngine
│   │   ├── engine.py                   # Concept graph, spreading activation, hippocampal indexing
│   │   └── __init__.py                 # + backward compat: ConceptGraph, ConceptNode, etc.
│   ├── decoder/                        # DecoderEngine
│   │   ├── engine.py                   # Neural decoder, vocab, training, generation
│   │   └── __init__.py
│   ├── web/                            # WebLearner
│   │   ├── learner.py                  # Multi-API search, background learning, curiosity
│   │   └── __init__.py
│   ├── bootstrap/                      # BootstrapManager
│   │   ├── manager.py                  # Consolidated: auto_expand, seed_from_curiosity, bootstrap_domain
│   │   └── __init__.py
│   ├── nn/rlm/                         # RLMv2 Decomposed
│   │   ├── relation_predictor.py       # RelationPredictor (W_rel, verb-stem offset)
│   │   ├── propagation.py              # PropagationEngine (multi-phase spread)
│   │   ├── plasticity.py               # Plasticity (Hebbian/Anti-Hebbian/Structural)
│   │   └── __init__.py
│   ├── chat/                           # ChatInterface
│   │   ├── interface.py                # CLI, response pipeline, reasoning loop
│   │   └── __init__.py
│   └── pyproject.toml                  # numpy>=1.20
├── ravana_ml/                          # ML Framework (~5,200 lines)
│   ├── nn/rlm_v2.py                    # RLMv2 (triple decomposition, GloVe)
│   ├── graph.py                        # ConceptGraph (3,678 lines)
│   ├── tensor.py                       # StateTensor / RawTensor / Parameter
│   ├── plasticity.py                   # Hebbian, Anti-Hebbian, Structural
│   └── ... (tokenizer, embedder, free_energy, currencies, propagation)
├── ravana-v2/                          # GRACE Cognitive Core (~19,500 lines)
│   ├── core/                           # 27 Phases A–P modules
│   │   ├── governor.py                 # Central regulation
│   │   ├── identity.py                 # Momentum-based self-concept
│   │   ├── sleep.py                    # 4-stage SWS+REM
│   │   ├── emotion.py                  # VAD differential equations
│   │   ├── human_memory.py             # Episodic/Semantic (2,321 lines)
│   │   ├── global_workspace.py         # Consciousness bottleneck
│   │   ├── meaning.py                  # Intrinsic motivation
│   │   ├── dual_process.py             # System 1 / System 2
│   │   └── ... (20 more modules)
│   ├── experiments/phases/             # Phase runners
│   └── tests/                          # Unit tests
├── experiments/                        # Cross-domain & Phase 4 experiments (18 files)
├── scripts/                            # Analysis & evaluation tools
│   ├── run_ablation.py                 # Ablation study runner (NEW)
│   ├── triple_eval.py                  # Per-triple evaluation harness (NEW)
│   └── ...
├── tests/                              # ML framework tests (17 files)
├── docs/                               # Full documentation (12 files)
├── data/glove/                         # GloVe embeddings (download required)
├── results/                            # Benchmark outputs & diagnostics
├── ablation_results.json               # Ablation study outputs
├── triple_eval_report.json             # Per-triple diagnostic reports
└── RAVANA_DEV_PLAN.md                 # Development roadmap

---

Ablation Study Framework (NEW)

The modular package includes a systematic ablation runner (scripts/run_ablation.py):

# Full factorial (4 flags × 3 modes = 48 configs)
python scripts/run_ablation.py --queries "hello|what is trust|explain oxiverse|bye" --runs 3

# Quick mode (baseline + 4 single removals + 2 mode variants = 7 configs)
python scripts/run_ablation.py --quick --modes

# Output: ablation_results.json + markdown comparison table

What it tests:

Flag	Component
--no-vad	VAD emotion modulation
--no-rlm	RLMv2 triple verification
--no-beliefs	Multi-user belief store
--no-curiosity	Autonomous curiosity-driven learning

Metrics tracked: Success rate, avg latency, total time, error modes per config.

---

Per-Triple Evaluation Harness (NEW)

Detailed diagnostics instead of averaged metrics (scripts/triple_eval.py):

# Default 10 triples (semantic, causal, contrastive)
python scripts/triple_eval.py --data-dir ./data --output triple_eval_report.json

# Custom triples
python scripts/triple_eval.py --triples-file my_triples.json --output report.json

Per-triple metrics:

• Relation type: semantic / causal / contrastive / possessive / temporal / analogical
• Prediction error (PE): how well the model predicts the object
• Confidence: model's confidence in the prediction
• Top-1 / Top-5 accuracy: exact match / in top 5
• Rank: position of correct object in predictions
• Graph edge: weight, confidence, type, prediction_free_energy
• Source: where learned — seed | web | user | sleep | inference | unknown

Aggregated diagnostics:

• Overall metrics + by relation type + by source
• Edge health summary (avg weight, confidence, PE, type distribution)

---

License

MIT — Built for the RAVANA-AGI-Research initiative.

---

About

RAVANA explores an alternative paradigm for machine learning — where cognition emerges from internal pressure, not gradient optimization; where knowledge self-organizes through Hebbian plasticity and sleep consolidation; and where identity, emotion, and meaning are first-class citizens of the learning dynamics.

The core thesis: "Cognition as pressure-driven self-organization."

This is the center of gravity. Everything else is implementation detail.

---

Citation

If you use RAVANA in research, please cite:

@misc{ravana2026,
  title={RAVANA: Pressure-Driven Self-Organization for Cognitive Machine Learning},
  author={RAVANA Research Team},
  year={2026},
  url={https://codeberg.org/oxiverse/ravana}
}

---

Links

• Documentation: docs/
• External Audit: docs/EXTERNAL_AUDIT.md