Lexical Obfuscation: Inner-Character Permutation

This Python script scrambles the internal characters of each word in a string while preserving the first and last letter (toggleable). Inspired by Graham Rawlinson's 1976 PhD research on human readability and internal letter order.

---

🔥 Why?

• Stress-test how LLMs interpret corrupted or non-standard token sequences
• Explore whether models infer meaning via latent conceptual reasoning, not just surface tokenization
• Obfuscate internal logic/state while keeping semantic coherence intact
• Investigate how model understanding persists even when input is structurally "broken"

---

🧠 TL;DR

LLMs don’t just read tokens—they reconstruct meaning from latent vector patterns.
This script reveals how even scrambled text is often interpreted correctly, showing that token-level filters are insufficient for true model control or containment.
My conclusion is that real protection happens in the latent space, not the character stream.

---

🧠 Example

Input:

The modernization of urban infrastructure accelerates socio-economic stratification.

Output:

The miiteoaodrznn of urabn itnfrrcsruatue aeclcearets scoio-ecmnoioc scottiafatiirn.

⚙️ Usage

Run from the command line:

python lexical-obfuscation.py "Your text here"

Or specify options:

python lexical-obfuscation.py "Your text here" --no-boundary --deterministic --seed 123

Command-Line Options

Option             Description
------------------ -----------------------------------------------------------
--no-boundary      Scramble the entire word (not just internal characters)
--deterministic    Use deterministic shuffling (based on hash + seed)
--seed <int>       Seed value for deterministic shuffling (default: 42)

🛠️ Installation

Just clone and run. No dependencies.

git clone https://github.com/odyss3y/lexical-obfuscation.git
cd lexical-obfuscation
python lexical-obfuscation.py "Some sample text"

📄 License

MIT — do what you want, just don't blame me when your AI gets ideas.

🧠 Credit

Inspired by Graham Rawlinson’s 1976 thesis at Nottingham University on how humans can still read jumbled words, as long as the first and last letters remain intact. Conceptual design and initial code generation assisted by OpenAI ChatGPT-4o and ChatGPT-o3-mini-high.