sealedstructs

Introduction

Experimental implementations of persistent (immutable) data structures in pure Python, using structural sharing and lexical encapsulation.

Contents

• ListNode -- persistent singly-linked list
• PersistentMap -- persistent hash map with 4-way hash trie

Rationale

Python relies on conventions (leading underscores, documentation, discipline) to enforce encapsulation. This works well when all code is trusted and carefully written but two trends are eroding that assumption:

1. Supply-chain risk: Malicious or compromised packages are increasingly common. The aiocpa incident is one example; the link notes PyPI receives over 500 malware reports per month.

2. High-volume generated code: As LLMs produce more code, the audit burden grows. Structural invariants -- enforced by construction -- reduce the surface area where subtle bugs can hide.

A particular concern with generated code is training data poisoning: if attackers introduce enough subtly flawed examples into training data, models learn to reproduce those patterns. The resulting bugs -- off-by-one errors, race conditions, type confusion -- look like ordinary mistakes but may be systematically exploitable. Unlike malicious packages, this vector doesn't require compromising any single dependency; the flaw arrives through otherwise legitimate tooling.

These structures take a different approach: internal state is sealed inside lexical closures, so there's no obj._foo to accidentally access, reassign, or corrupt -- the state does not exist as attributes. Structural invariants enforced by construction reduce the surface area where subtle bugs -- whether accidental or induced -- can hide.

Limitations

This is defense in depth, not a security boundary. Python's runtime allows introspection and mutation of closure internals: __closure__[0].cell_contents is writable. Malicious code running in the same process can bypass any pure-Python encapsulation.

These structures protect against accidental misuse (including from sloppy dependencies or generated code). They do not protect against intentional attacks. True isolation requires process boundaries, containers, or a different runtime.

Development

python -m venv .venv
source .venv/bin/activate
pip install -e .
pytest

Or to add to your project:

pip install git+https://github.com/aschweig/sealedstructs.git