Story

Hi HN — I’ve been working on a small open-source experiment around runtime AI safety.

The idea is to treat AI risk not as a binary “safe/unsafe” state or post-hoc failure analysis, but as accumulated constraint strain over time — similar to how engineers think about mechanical stress.

The project defines a simple, model-agnostic signal (GV) and a lightweight runtime monitor (Sentinel) that classifies risk into bands (green/yellow/red) and suggests interventions (alert, throttle, human review).

This is an early MVP — intentionally minimal — meant to explore whether continuous, quantitative safety signals are useful before failures occur, especially for agents and LLM-based systems in production.

I’d really appreciate feedback, criticism, or pointers to prior art I should study. Repo: https://github.com/willshacklett/gvai-safety-systems