Reality-Grounded Autonomous Intelligence: An Evidence-, Authority-, and Outcome-Governed Architecture for Reliable AI Agents

@misc{rm3692026reality,
  author       = {{RM369}},
  title        = {{Reality-Grounded Autonomous Intelligence: An Evidence-, Authority-, and Outcome-Governed Architecture for Reliable AI Agents}},
  year         = {2026},
  month        = oct,
  date         = {2026-10-30},
  publisher    = {Life of Research},
  howpublished = {Preprint},
  url          = {https://lifeofresearch.com/preprints/164},
  abstract     = {Autonomous AI agents increasingly combine language-model reasoning with tools that can inspect, modify, and act upon external environments. Their reliability therefore depends on more than the quality of generated text: an agent must preserve the user's authorized objective, distinguish observations from beliefs and predictions, operate within enforceable permissions, account for partial and irreversible effects, and establish whether the intended outcome actually occurred. Existing techniques address important pieces of this problem—including reasoning-action loops, tool-mediated critique, provenance, risk management, and transactional recovery—but these mechanisms are often deployed without a single operational model that governs the transitions between intention, knowledge, authority, execution, and acceptance.

This paper proposes Reality-Grounded Autonomous Intelligence (RGAI), a system architecture for governing tool-using AI agents. RGAI separates five responsibilities: objective governance, epistemic integrity, authority governance, transactional execution and recovery, and independent assurance. These responsibilities share a versioned event history from which current state views are derived. RGAI distinguishes user-confirmed requirements from preferences and predictions; observations from external assertions and model inferences; permission to prepare from permission to commit; an attempted action from a confirmed effect; and a confirmed effect from an accepted outcome. A small trusted enforcement core mediates consequential actions, while risk-adaptive assurance allocates verification effort according to potential harm, uncertainty, irreversibility, and blast radius. The architecture also addresses prompt injection, privacy leakage, stale permissions, hidden dependencies, multi-agent concurrency, poisoned memory, reward tampering, and ambiguous external outcomes.

RGAI is a proposal, not a demonstrated improvement. The paper specifies its state model, action contract, transition rules, trust assumptions, failure handling, and a matched evaluation programme. It proposes testing against competent conventional agents on established computer-use, web, multi-application, research, and tool-interaction benchmarks augmented with controlled adversarial conditions. The central hypothesis is that explicit governance of state transitions can improve safe task completion and failure containment without imposing prohibitive costs on ordinary tasks. Whether that benefit exists, and at what operating point, remains an empirical question.},
  keywords     = {autonomous agents, AI reliability, tool use, evidence provenance, action authorization, runtime assurance, transaction recovery, agent security},
  language     = {en},
  copyright    = {Creative Commons Attribution 4.0}
}