pklm-sandbox: Deterministic Token Masking for Offline Edge AI

#devchallenge #hf26challenge

Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass Submission 🌿

*This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass*

What I Built

I built pklm-sandbox, an open-source testing repository showcasing token-level logit masking (PKLMSandboxProcessor) designed to eliminate stochastic drift and ensure strict structural safety.

For the "Touch Grass" theme, this project serves as a deterministic safety and constraint middleware for offline field-assistance tools (such as offline plant/animal identification or wilderness trail assistants running on edge hardware). When out in the field with no signal, users need absolute guarantees that a local open-weight model will not hallucinate dangerous advice (e.g., misidentifying toxic flora) or output malformed data. pklm-sandbox enforces hard neuro-symbolic constraints directly at the logit level, ensuring the model's outputs remain structurally valid and safe.

Demo

Hugging Face Show & Tell Discussion

Code

GitHub Repository: TheImmortalPython/pklm-sandbox

pip install git+https://github.com/TheImmortalPython/pklm-sandbox.git

How I Built It

I used an open-weight local model architecture combined with Hugging Face pipelines, integrating a custom LogitsProcessor (PKLMSandboxProcessor) written in Python.

Instead of relying on post-hoc parsing or hoping a prompt injection will keep a model in line, the project is built around direct logit-level manipulation—masking out unauthorized token probabilities before the generation step even occurs to guarantee absolute determinism and compliance.

Why Does Open Innovation Matter?

Open innovation and open-weight models are vital here because edge-deployed field tools require complete control over the underlying weights and execution pipeline. A closed, proprietary API does not allow you to intercept token logits, modify probability distributions, or run entirely offline on local hardware out on a trail. Open-source infrastructure makes it possible to build transparent, predictable, and privacy-first AI systems that you can actually trust when you are miles away from an internet connection.