It computes a bounded “GV” signal from live agent behavior (token velocity, tool calls, errors, recursion, repetition) and emits: green / yellow / red → continue / slow / halt.

The goal is runtime survivability rather than training-time alignment. No model introspection, deterministic scoring, and designed to be embedded directly into agent loops.

There’s a short demo script in the repo that simulates an agent going unstable. Feedback welcome.