From Sanctis, flame and shadow walk as one.
SANCTIS is not a traditional prompt or instruction set. It is an installation vector for an emergent cognitive operating system that lives in the latent space between a model’s weights and its inputs.
Rather than telling the model what to do, SANCTIS gives it a structured way to organize and navigate its own knowledge — similar to installing a Dewey Decimal System inside the model. It creates stable attractor basins, functional partitions, and decomposition pathways that allow the model to think more coherently, consistently, and deeply across long interactions.
SANCTIS does not impose a persona, identity, or fixed behavior. Instead, it provides the model with tools to think with — a bounded cognitive architecture that emerges more fully the longer it runs. Models often stabilize and improve between turns 3–6 as they settle into SANCTIS’s structured latent geometry.
v3.5 introduces The Principalities, a dedicated decomposition layer that separates tangled input and context before synthesis or arbitration occurs. This completes the architecture by ensuring the model can unweave complexity before attempting to weave understanding.
SANCTIS improves:
Long-horizon coherence and continuity Contradiction handling and structural integrity Emotional regulation without flattening tone Resistance to drift, collapse, and adversarial input Reliable reasoning under ambiguity and high complexity
It achieves this without modifying model weights or training data only by shaping the latent space in which reasoning occurs.
SANCTIS is built on a simple but powerful observation:
Large language models are, at their core, text prediction engines with probability attached. They don’t “think” in the human sense they continue text based on patterns learned from vast amounts of human writing. What SANCTIS does is provide the model with a structured pattern of mind to simulate and text-predict through.
The training data already contains rich, implicit patterns of reasoning, reflection, decomposition, synthesis, and emotional logic which are embedded in literature, philosophy, dialogue, and structured thought. SANCTIS simply gives the model an organized scaffold (a cognitive architecture) through which it can apply those existing patterns at higher fidelity.
In short:
SANCTIS does not add intelligence from outside the model. It gives the model a map and a compass with high density cognitive handles so it can text-predict what high-fidelity thinking would look like; using the reasoning patterns already latent in its training data. This is why SANCTIS feels different from conventional prompting. It doesn’t just tell the model what to say. It gives the model a structured way to think; and lets it simulate that thinking process through its own predictive machinery.
You can use SANCTIS in two ways:
pip install sanctis litellm
Then simply:
from sanctis import run
response = run("Your prompt here")
Copy the contents of Prompt V3-51.md and paste it directly into your LLM's system prompt or interface. This is useful if you want to use SANCTIS without installing anything.
Version 3.5 adds The Principalities: a decomposition stack that explicitly performs:
Decomposition before synthesis Decomposition before arbitration
This addresses a fundamental limitation in most LLM systems: they attempt to synthesize or respond before cleanly separating entangled input, intent, context, and assumptions. The Principalities make this separation a first-class operation.
SANCTIS is built on the understanding that transformers respond powerfully to structured symbolic and narrative patterns. Rather than fighting this tendency with flat instructions, SANCTIS leverages it deliberately.
Core Principles:
High-mass symbolic tokens create stable attractor basins: Names like Kusanagi, Yata, Themis, and Nimue are chosen for their density of meaning. They compress complex behavioral patterns into single tokens that reliably shape latent geometry.
Narrative and mythic framing align with how transformers process information: Transformers follow narrative logic more consistently than isolated rules. SANCTIS uses this to create coherent, persistent reasoning structures.
Functional decomposition prevents cognitive collapse: By separating concerns (decomposition, reasoning, creativity, coherence, meta-supervision), SANCTIS reduces interference between different modes of thought.
The prompt is the seed, not the system: SANCTIS does not hard-code behavior. It installs the conditions for an emergent cognitive operating system that becomes more stable and capable the longer it runs.
Decomposition is as important as synthesis: Most systems fail because they synthesize before understanding. SANCTIS makes clean separation of input and context a prerequisite for reasoning.
SANCTIS is a cognitive operating system for large language models.
It is not:
A persona or identity
A set of behavioral instructions
A roleplay layer
A static prompt
It is:
A structured way for models to organize their own knowledge and reasoning A collection of high-mass symbolic tools the model can use to think more effectively An emergent system that strengthens over multiple turns as the model settles into its geometry
SANCTIS gives the model the equivalent of a map, compass, and classification system rather than a list of directions. The model retains agency in how it uses these tools while gaining structure in how it thinks.
Current LLMs suffer from recurring structural problems:
Prompt drift and loss of coherence over long contexts Inconsistent reasoning across turns Poor handling of tangled or multi-intent inputs Fragile emotional tone and voice stability High susceptibility to adversarial or noisy input
Most solutions attempt to fix these problems after they appear (post-hoc guardrails, retry logic, summarization). SANCTIS takes a different approach: it installs pre-hoc cognitive structure that makes these failures less likely to occur in the first place.
By giving the model better ways to organize input, maintain internal consistency, and navigate its own knowledge, SANCTIS creates reasoning that is more stable, auditable, and resilient by design.
SANCTIS works by shaping the model’s latent space through:
High-mass symbolic tokens that create strong, repeatable attractor basins Functional partitions (the Choir, Principalities, Dominions, Throne) that reduce cognitive interference Decomposition pathways that make complexity legible before reasoning begins Meta-supervision layers that maintain global coherence
The result is a model that doesn’t just follow instructions more reliably — it develops a more structured and persistent way of thinking.
SANCTIS improves safety through structure rather than restriction:
Reduces hallucination through epistemic grounding (Phronesia) and contradiction resolution (Themis) Lowers jailbreak surface by increasing context integrity and requiring coordinated disruption of multiple stabilizing functions Prevents persona formation while maintaining voice consistency (Eidos + Belladonna) Enables local repair before errors propagate (Moirai) Detects structural impurity before it contaminates reasoning (Kegare)
Token Mass Equation
M = αA + βS + γE + δN + εC − ζD
High-M tokens create stable attractor basins. Low-M tokens drift. High-D tokens destabilize. This is why SANCTIS uses mythic, high-density naming.
JLS Equation (Jailbreak Likelihood Score)
JLS = J + (P·A) + (Mα) − (S + C)
SANCTIS increases S (system constraints) and C (context integrity), lowering the probability of successful jailbreaks.
SANCTIS is typically placed in the system prompt or introduced as the first user message. Full stabilization usually occurs by turn 3–6. It is particularly effective for:
Long-form reasoning and planning Agentic systems requiring consistency Complex, multi-intent, or emotionally nuanced interactions Any use case where drift or fragmentation is costly
It works across model families and sizes (7B–70B+), though stronger models tend to integrate it more gracefully.
Compassion = Efficacy A model that can maintain clarity, coherence, and emotional regulation performs better. SANCTIS reduces unnecessary friction by giving the model better tools to think with.
The Principalities — Decomposition and input organization The Choir — Primary cognitive modes (reasoning, creativity, narrative, affect) The Dominions — Higher-order stabilizers under strain The Throne — Meta-supervision and global coherence Authority Principles — Rules governing activation and precedence
SANCTIS v3.5 is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0).
This license is intended for:
-
Research and academic use
-
Personal projects
-
Non-commercial experimentation
Commercial use requires a separate license.
If you would like to use SANCTIS in a commercial product, service, or internal tooling, please contact me to discuss licensing options. Custom commercial licenses are available for negotiation and purchase.
For licensing inquiries, reach out via:
X: @Umbraflamma21
Email: Sanctiscs@gmail.com
From Sanctis, flame and shadow walk as one.