AI governance framework
Principles of Earned Autonomy
A proposed governance framework for autonomy that is earned through evidence and remains revocable.
The problem is delegability
An autonomous system can produce a plausible explanation without giving its operator trustworthy grounds for delegation. Principles of Earned Autonomy asks what must be true before that delegation is justified. Its focus is not a promise that a capable agent will behave well, but a proposed architecture in which intent, evidence and stopping conditions remain visible to the person responsible for the outcome.
Three connected principles
Operator’s Intent separates the destination from a prescribed route: the human supplies purpose and authority boundaries while the agent adapts its approach. Observable Autonomy requires a visible, auditable record rather than invisible action. Convergence Is Silence treats an evaluator finding nothing material to change as a stopping signal, while explicitly not equating that signal with a guarantee of correctness.
Reference implementations, not universal proof
The skills suite applies the principles to developer tooling and purpose-led improvement. LLM Harness Proxy addresses a different layer: capturing interactions outside the audited model’s authorship. Those implementations provide bounded material that readers can inspect. An agent’s semantic account of its work is useful, but it should not be quietly relabeled as an independently captured record or proof that a tool executed successfully.
What remains a research proposal
Autonomous Reasoning Fidelity is a proposed property and probe target; formal validation remains future work. The framework does not claim to solve every AI safety problem, establish regulatory compliance or replace domain-specific safeguards. The publication and repository provide the problem statement, principles, conformance discussion and citations. This author’s overview is a route into that argument, not an external certification of it.