What this is. A proposed theory-level governance framework for how autonomous AI agents may earn and exercise authority. The claimed contribution. Not the isolated claim that reasoning matters, nor the premise that self-narration is unreliable. The claimed contribution is the synthesis: two core problems (autonomous reasoning and earned autonomy), three architectural principles (Commander's Intent, Observable Autonomy, Convergence Is Silence), and one proposed measurable property (Autonomous Reasoning Fidelity, ARF) combined into a discipline of earned, observable, revocable authority. Why it matters. The question is not only whether an agent can produce useful outputs, but on what evidence a human or institution should let it act more autonomously in a specific context. Scope posture. This is presented as a step toward deployable governance for delegability, not as a complete solution to AI safety or autonomy. Evidence status. Public conformance evidence lives in the released documents and the two separately published reference implementations they cite: the Principles of Earned Autonomy Skills Suite (developer-tooling domain, three-family silence-convergence run) and the LLM Harness Protocol (a transparent MITM proxy that writes a tamper-evident, hash-chained, append-only ledger of every LLM interaction across OpenAI, Anthropic, and Gemini APIs, demonstrating that the structural capture-author separation Principle 2 requires is buildable in current tooling). Core argumentative line. Problem -> principles -> bounded reference evidence from two implementations. Read in this order. README.md - overview and scope. PROBLEM.md - names the two problems and defines delegability as the connecting discipline. PRINCIPLES.md - states the three principles and the ARF operational definition. PROOF.md - conformance tests for each principle, plus bounded empirical evidence from the two reference implementations. Background and corroboration. EMPIRICAL_EVIDENCE.md records formative case material from the framework's synthesis plus external corroboration for why the principles are structural; it is not part of the core evidentiary line. Canonical source: GitHub repository. Companion implementations: Autonomous Development Skills Suite on Zenodo and the LLM Harness Protocol.