Welcome — run the non-bypass proof yourself, and try to break the gate
Hi all, Author here. This is the official discussion thread for Antaḥkaraṇa, a governed runtime for continual-learning agents: one audited control ring between a model's proposals and any real-world action, learning that can only make the gate stricter on its own, and a machine-checked proof that nothing bypasses the ring.
Try it in two minutes:
pip install antahkarana
atk verify --teeth
--teeth first proves the non-bypass invariant over the full reachable state space, then deliberately breaks the model three ways and shows you the checker catching each one with the shortest counterexample. Every number in the model card comes from the gate runs (G7–G12) in the repo — nothing is illustrative.
What I'd love from this community:
- Break it. If you can get a policy escalation or an ungated action past the ring in the shipped SDK, I want to know, you'll be credited in the model card.
- Tell me your use case. What would you point a governed loop at? (Ops monitoring, procurement watching, agent guardrails…) Real workloads shape the roadmap.
- Challenge the claims. The proof's scope is stated precisely (bounded model of the coordination skeleton); if you think the scope is too narrow or the framing overstates it, argue it here.
Questions, bug reports, and skepticism all welcome — skepticism especially. That's what the gates are for.