SANCTIS v3.51

A cognitive architecture that gives LLMs structured thinking patterns — without the copy-paste.

What is SANCTIS?

SANCTIS is a model-facing cognitive architecture that provides LLMs with structured thinking patterns (decomposition, synthesis, arbitration, reflection, etc.) so they can reason more coherently, maintain context better, and behave more consistently over long interactions.

Think of it as giving the model a "cognitive operating system" — it doesn't make the model smarter, but it gives it better tools to think with.

Key insight: LLMs are already capable of sophisticated reasoning. SANCTIS simply gives them a reliable structure to use those capabilities more effectively.

Why SANCTIS?

Most LLM failures aren't intelligence failures — they're coherence, continuity, and behavioral stability failures.

SANCTIS helps with:

• Long-horizon coherence
• Reduced contradiction and drift
• Better multi-turn conversations
• More consistent agent behavior
• Structured reasoning without manual prompting

Installation

pip install sanctis litellm
That's it. No more copying 390 lines of prompt.

Quick Start

Basic Usage

Python

from sanctis import run

response = run("Explain quantum computing simply")
print(response)

Multi-turn Conversations

Python

from sanctis import Conversation

conv = Conversation()
response1 = conv.run("My name is Luke.")
response2 = conv.run("What is my name?")
print(response2)  # "Your name is Luke."
With LangChain

Python

from sanctis.langchain import get_sanctis_runnable

llm = get_sanctis_runnable("ollama/llama3")
response = llm.invoke("Tell me a story about a robot learning to feel.")
print(response.content)

Command Line

Bash
sanctis run "Explain how transformers work"
sanctis setup

Key Features

Smart Defaults — Automatically detects Ollama or OpenAI
Configuration System — Set defaults once (sanctis config set model ollama/llama3)
Generation Parameters — Control temperature, max_tokens, top_p, etc.
Streaming Support — Real-time responses in CLI and Python(LangChain support coming soon)
LangChain Integration — Drop-in Runnable for agents and chains
Professional CLI — Includes interactive setup wizard
Clear Error Messages — Helpful guidance when things go wrong


How It Works (Simple Explanation)
LLMs are fundamentally text prediction engines. SANCTIS gives them a structured pattern of thinking (inspired by how humans actually reason) that they can "text predict" their way through.
Instead of hoping the model figures out good reasoning on its own, SANCTIS provides:

Decomposition modes
Synthesis modes
Arbitration & contradiction resolution
Reflection & refinement modes
Narrative & creative modes

The model activates these as needed — silently, unless you ask it to show its work.

Configuration

Bash

# Set a default model
sanctis config set model ollama/llama3

# List all settings
sanctis config list

# Get a specific setting
sanctis config get temperature

Advanced Usage
Custom Generation Parameters
Python

from sanctis import run, Conversation

# Single turn with custom settings
response = run(
    "Write a detailed technical explanation",
    temperature=0.3,
    max_tokens=2000,
    top_p=0.9
)

# Conversation with custom settings
conv = Conversation(temperature=0.8, max_tokens=500)

Streaming

Python

from sanctis import run

sfor chunk in run("Tell me a long story", stream=True):
    print(chunk.choices[0].delta.content, end="", flush=True)

Philosophy

"Compassion = Efficacy"

Clarity, coherence, and stability increase performance and reduce friction. SANCTIS is built on the belief that giving models better structure to think with leads to better outcomes than trying to force better behavior through constraints alone.

License
SANCTIS is free for research and evaluation.
Commercial deployment requires a licensing agreement.
© Luke "Umbraflamma" Yuro 2025–2026
From Sanctis, flame and shadow walk as one.