I built Nobulex to fix this. It is open-source middleware that does three things:

1. Lets you define behavioral rules in a simple DSL (permit, forbid, require)

2. Intercepts all actions at runtime and blocks anything that will violate the rules

3. Logs everything in a hash-chained audit trail that anyone can independently verify, not just the operator.

The key insight: you can't just audit a neural network's reasoning, but what you can do is audit its actions against stated commitments. `verify(rules, actionLog)` is always deterministic.

    npm install @nobulex/identity @nobulex/covenant-lang @nobulex/middleware

    covenant MyAgent {
      permit read;
      forbid transfer where amount > 500;
      require log_all;
    }

I'd love feedback on the rule language — is the permit/forbid syntax intuitive, or would you design the DSL differently?

I'm 15 and built this solo. Happy to answer anything about the architecture.