Verification before autonomy
Autonomy is not a slider you turn up. It is something you earn, one verified decision type at a time.
The temptation with a capable model is to hand it the keys: let it read, decide, and write in one motion. That works until the first decision you can’t explain, and then you spend a fortnight reconstructing what happened from logs that weren’t designed to answer the question.
Invert the default
We made refusal the default. An agent does nothing irreversible unless a check it cannot author itself has passed. The model writes a proposal; a separate, deterministic verifier decides whether the proposal is allowed to act. The two are never the same code path.
The review queue is the product
The queue where refused proposals land is not a failure state — it is the safest, most informative surface in the system. It is where the model tells you it is out of its depth, and it is the dataset you use to decide what to automate next.