Getting Started — Character Identity Protocol (CIP)

Character Identity Protocol (CIP) is an operational governance protocol for stabilizing character identity in probabilistic generative systems. It controls the conditions under which identity convergence occurs — not by modifying the model itself, but by governing how generation is managed.

This page guides you through the key concepts before introducing the protocol.

The Problem

Generative AI models often fail to reproduce the same character consistently across generations.

Even when prompts remain identical, the generated character may change.

This happens because:

generative models reconstruct outputs probabilistically, not deterministically
prompts are internally reinterpreted before generation (A → A′)
small changes in context, pose, or wording can shift the reconstruction result

This phenomenon is known as identity drift.

→ See: Character Drift Taxonomy

How Generative AI Actually Behaves

AI systems do not execute prompts directly.

Instead, they internally reinterpret inputs before generating outputs.

Conceptually:

A (user input)
→ A′ (internal interpretation)
→ B′ (generated output)

A′ is shaped by internal constraints (C) — including training priors, optimization pressure, and compression — which is why B′ may deviate from what A intended.

→ See: Core Model — A → (A + C) → B′

Because generation is probabilistic, identical inputs may not produce identical outputs. Results can vary between generations and drift over time.

→ See: How Generative AI Actually Behaves

Prompt Structure

Prompts work best when information is structured and concise.

A simple structure for general prompts:

Element	Role
Theme	What the prompt is about
Rules	How the AI should respond
Materials	What information it should use
Output	How the answer should appear

→ See: A Simple Structure for Writing Prompts

Writing Prompts for Image Generation

Image generation prompts typically contain four types of signals:

Element	Role
Subject	Main object or scene
Style	Visual rendering style
Details	Additional visual elements
Constraints	Restrictions or exclusions

For character generation, placing identity-related information early in the prompt helps maintain consistency.

→ See: Writing Prompts for Image Generation

The Character Consistency Problem

Even with structured prompts, the same character may appear differently across generations.

Common drift types include:

Identity Drift — the character becomes a different person
Style Drift — rendering shifts toward photorealism or a different style
Attribute Drift — character attributes gradually disappear from the output

These drift types arise from the probabilistic nature of generative models and the way prompts are internally compressed.

→ See: Character Drift Taxonomy

Want to see this for yourself? Run the experiment described in:

→ Why Your Character Keeps Changing — experiment-first explanation with a runnable A/B workflow
→ Character Consistency — An Observation — observation-based companion article

Character Identity Protocol (CIP)

CIP is a structured workflow designed to manage identity convergence.

It operates through four elements:

Element	Role
Anchor Image	Previously validated identity reference
Minimal Prompt	Lightweight identity constraints
Identity Gates	Validation checks (Face ∧ Skeleton ∧ Proportion)
Hard Abort	Immediate termination when drift is detected

CIP does not modify the model. It governs the operational conditions under which generation is allowed to continue.

→ See: White Paper — Technical Mechanism

Quick Start

Ready to try CIP?

→ Quickstart Guide

Documentation

Concepts

The Problem

Protocol

Guides

Reference