Evidence-bound autonomy is an architecture for AI agents in which no consequential action executes until the agent has produced its evidence: the signals it read, the policy it applied, the reasoning it followed, and the reversibility it can guarantee — sealed into a tamper-evident record before the action, not after.
The inversion is the point. Conventional automation acts and then logs; if the log is incomplete, the organisation discovers it during an incident review. An evidence-bound agent cannot act without its proof packet, so the record is complete by construction.
Under this model the distinction between automated and manual work stops being a risk category: both produce the same artefact — a decision with its evidence attached, attributable to a policy and a person. Auditors review decisions, not systems.
Evidence-bound autonomy is the practical answer to the governance question boards actually ask about agentic AI: when this goes wrong, will we be able to show we were in control?
This vocabulary is implemented, not aspirational — it describes how the MAIA substrate actually runs. See the platform or read the research.