I've open-sourced Invariant Governance, a framework for enforcing authorization boundaries on autonomous agents, trading algorithms, and robotic systems without adding human-approval latency. The core problem: existing governance forces a choice between human-in-the-loop (safe but slow) and probabilistic guardrails (fast but brittle). LLM-based filters can be gamed. Confidence thresholds can't detect salami-slicing where each action is within bounds but the cumulative effect is catastrophic.

  The framework separates authority from execution from observation using three structurally decoupled components. The
  Governance Kernel evaluates policy and issues cryptographic approval receipts but cannot execute actions. The Execution
  Gate validates receipts but cannot authorize itself. The Telemetry Observer records everything through a one-way channel
   and cannot intervene.

  No ML. No probabilistic scoring. Deterministic enforcement only.

  Apache 2.0 with defensive patent grant. Open-core model.

  GitHub: https://github.com/utahbroker/invariant-governance

  Docs: https://invariant-governance.com

  Feedback welcome, especially from folks building production autonomous systems.