A working definition
A system operates with autonomy when it can complete a decision loop without a human in the middle of every step. It operates with evidence when it can reconstruct the context, policy, alternatives, actor and effect behind any action it took.
Autonomy without evidence is an incident waiting for an owner. Evidence without autonomy is a very expensive log file.
The four controls
Policy boundaries. The system is given an explicit envelope of what it may do, to what, at what value, at what frequency. Anything outside the envelope escalates rather than improvises.
Graduated approval. New action types begin as recommendations, move to approve-before-execute, then to execute-with-notification, and only then to autonomous. Movement between stages is earned with measured outcomes, not asserted in a roadmap.
Traceability. Every action carries the reasoning, inputs, policy version and actor identity that produced it, stored in a way that survives the session and can be handed to an auditor without a translation layer.
Reversibility. Where an action can be undone, the undo path is built at the same time as the action. Where it cannot, that irreversibility is treated as a design constraint that raises the approval bar.
The failure modes we design against
We design against four failure modes: silent drift, where a system keeps acting correctly against a world that changed; confident scope creep, where a loop that worked on easy cases quietly starts handling hard ones; accountability voids, where nobody can name the human who owns a decision; and evidence theatre, where the logs exist but couldn't answer a real question.
“Autonomy is not a switch you flip. It is a budget you spend, and evidence is the only currency that tops it back up.”