Skip to content

Sovereign Cognition & Persona Architecture

The foundational premise of Tur is that "Prompt Engineering" is a dead end.

Prompt Engineering treats the interaction with an LLM as a literary exercise. Developers write paragraphs of prose, begging the model to act a certain way, follow specific rules, or avoid certain pitfalls. Because LLMs are probabilistic, this unstructured text often leads to unpredictable, non-deterministic, and fragile behavior. If the prompt is too long, the model suffers from "lost in the middle" syndrome. If the prompt is too vague, the model hallucinates.

Tur replaces literary Prompt Engineering with Sovereign Cognition & Persona Architecture.

Identity as Code (DNA)

In Tur, an AI persona is not an ephemeral text prompt; it is a structured, immutable software object and an evolving cognitive entity.

We define the persona using strict data schemas (Pydantic models serialized to clean state files). An identity consists of:

  • The Aleph: The core, unyielding motivation.
  • The Council: Distinct, weighted philosophical constraints (Policy layer).
  • Protocols: Trigger-action behavioral loops (e.g. Dennis Protocol, Speech Center).
  • Merkle Memory: Fractal L1/L2 memory graph representing the continuity of self.

By treating identity as sovereign code, we can version it, clone it, test it, and inject it predictably into the context window. We stop writing prose and start engineering cognitive topology.

Constraint Dimensionality (\(C_p\))

When you impose a constraint on an LLM (e.g., "Do not hallucinate," or "Always explain your reasoning"), you are physically altering the probabilistic landscape of the model's inference path. You are making the path more "rugged."

We call this Constraint Dimensionality (\(C_p\)).

Every principle and protocol added to a persona increases its \(C_p\). A higher \(C_p\) means the model is more strictly bound, but it also means it requires more compute (and a more capable underlying model) to successfully navigate the rugged landscape without degrading into gibberish.

The Telemetry Protocol

Tur includes a native telemetry command that calculates the \(C_p\) of your active persona.

  • Human (Manageable): Low \(C_p\). The persona is lightly constrained and can run on fast, smaller models (e.g., Gemini Flash, Claude Haiku).
  • Giant (Heavy Load): Medium \(C_p\). The persona is rigorously structured. Requires frontier models (e.g., Claude Opus, GPT-4o).
  • Titan (Inference Warning): High \(C_p\). The persona is over-constrained. The prompt has become so heavy with rules that the model will likely suffer from cognitive collapse or severe context entropy.

Sovereign state modeling is the discipline of balancing \(C_p\). You must provide enough structure to make the model safe and deterministic (The Golem Protocol), but not so much structure that it paralyzes the inference engine (The Shannon Module).