FerrellSyntheticIntelligence/Vitalis_Core

Vitalis Core

Blank slate, bolt-on cognitive framework for any model.

Vitalis Core is not a model. It has no trained weights, no training data, and no pre-loaded knowledge. It is a cognitive layer that bolts onto any model — HuggingFace transformers, llama.cpp, OpenAI API, or your own custom inference — to add:

• Memory — working, long-term, and context memory that persists across sessions
• Validation — classifies input/output as good or garbage before/after inference
• Sandboxing — tests every response in isolation before delivering it
• Truth checking — detects hallucinations, contradictions, and hedging
• Experimental flagging — marks low-confidence or uncertain outputs with [EXPERIMENTAL]

Usage

git clone https://huggingface.co/FerrellSyntheticIntelligence/Vitalis_Core
cd Vitalis_Core
pip install -e .

from vitalis import CognitiveLayer

# Any callable model
def my_model(prompt):
    return "model response..."

# Bolt Vitalis onto it
brain = CognitiveLayer(model=my_model)
result = brain.process("What is synthetic intelligence?")

print(result["response"])
# confidence=0.95, truthful=True, sandbox_passed=True

Architecture

User Input
   │
   ▼
┌─────────────────┐
│ MemoryEnrich     │  ← Attach working memory + LT recall
└────────┬────────┘
         ▼
┌─────────────────┐
│ DataValidator    │  ← Reject garbage/empty/abusive input
└────────┬────────┘
         ▼
┌─────────────────┐
│ Model (bolted)   │  ← Your model runs here
└────────┬────────┘
         ▼
┌─────────────────┐
│ ResponseSandbox  │  ← Safety, coherence, format check
└────────┬────────┘
         ▼
┌─────────────────┐
│ TruthEngine      │  ← Hallucination/contradiction detection
└────────┬────────┘
         ▼
┌─────────────────┐
│ ExperimentalFlag │  ← Tag low-confidence outputs
└────────┬────────┘
         ▼
   Final Output

Each stage is independently configurable via environment variables:

Variable	Default	Description
VITALIS_MEMORY_ENABLED	true	Memory enrichment
VITALIS_VALIDATOR_ENABLED	true	Input/output validation
VITALIS_SANDBOX_ENABLED	true	Response sandboxing
VITALIS_TRUTH_ENABLED	true	Truth/consistency checking
VITALIS_FLAGS_ENABLED	true	Experimental output flagging
VITALIS_DATA_DIR	./storage/	Data persistence directory
VITALIS_WORKING_MEMORY	20	Working memory buffer size

Examples

# Bolt onto any HuggingFace model
from transformers import pipeline
from vitalis import CognitiveLayer

pipe = pipeline("text-generation", model="microsoft/phi-2")
brain = CognitiveLayer(pipe)
result = brain.process("Explain gravity in one sentence.")

# Bolt onto llama.cpp
from vitalis import CognitiveLayer

def llm(prompt):
    return call_llama_cpp(prompt)

brain = CognitiveLayer(llm)
result = brain.process("What is the capital of France?")

See examples/ for complete working examples.

Zero Training

Vitalis Core ships with zero trained weights. The cognitive layer is pure logic:

• No pre-training data
• No fine-tuned parameters
• No external API dependencies
• No cloud requirement

It is a blank slate — ready to be bolted onto any model and customized for any domain.

License

GPL-3.0 — FerrellSyntheticIntelligence