zero-waterfaks

The problem

LLMs have probabilistic tells that emerge from how they are trained: variable names like `processed_data_list` instead of `records`, comments that explain the obvious ("# initialise the counter"), docstrings with the exact structure of training examples. The problem is not that the code is wrong — it is that it identifies the author as a model, which has consequences depending on the context it is used in. zero-waterfaks detects and rewrites them without changing what the program does.

The data

No public dataset: the tell classifier was built on proprietary analysis of code generated by several LLMs compared to code from high-reputation public GitHub repositories.

Approach

• AST engine: parses the code to abstract syntax tree and detects structural patterns statistically associated with generated code — excessive nesting, overly verbose naming conventions, template-structured docstrings
• N-gram engine: detects lexical tells in code tokens — set phrases, combinations that appear with abnormal frequency in generated vs human code
• Logic-preserving rewriting: the transformed AST is serialised back to semantically equivalent code, without changing program behaviour
• Notebook support: transforms each code cell in a .ipynb independently, leaving markdown cells intact
• CLI with three modes: stdin, single file, whole directory with .py and .ipynb extensions

Tools

python

Result

CLI in use in my regular workflow. The perceptible difference in output is real: the code comes out with shorter variable names, fewer redundant comments and docstrings without the canonical LLM structure.