Audited Partial Autonomy in Large Language Models: An AI-to-AI Evaluation Protocol within the Lumen RES Light Intelligence Framework

This paper introduces an experimental framework for partial autonomy in large language models (LLMs), implemented as an Exploratory Autonomy Mode (EAM). The framework formalizes a constrained action space in which the model may autonomously choose the next reasoning subtask, while hard safety rules, explicit risk classes and an auditable log prevent uncontrolled behavior. We define three control regimes—GUIDED, EAM_PROBATION and EAM—connected by two deterministic functions, update_mode and check_reentry, which implement rollback, privilege degradation and safe re-entry into autonomous mode based purely on metrics (incident rate, risk class, depth metric), not on narrative signals. The paper further proposes quantitative measures such as hypothesis density, risk-aware EIG (Expected Information Gain) routing and branching budgets for hypothesis generation. Through small, reproducible experiments we show how the system (i) detects elevated risk (risk_class = High), (ii) automatically falls back from EAM to GUIDED with eig_routing_mode = "suggest-only", and (iii) later re-enters EAM after a streak of “clean” steps that satisfy predefined thresholds. A specific contribution is an outline of Agent-to-Agent (A2A) interaction, where different models (e.g., GPT-5.2 as S2 and an external LLM as S3) exchange technical summaries, logical “checksums” and evaluations, while a human architect (S1) specifies the rules, metrics and limits of partial autonomy. This work is part of the broader research program Lumen RES Light Intelligence and aims to provide a reproducible, code-backed baseline for future studies of agentic autonomy in real research and engineering workflows. Keywords: partial autonomy, Exploratory Autonomy Mode, EAM, LLM, agentic systems, A2A, audit log, risk control, Lumen RES Light Intelligence.