AuditTune campaign evidence: experiment ledger, agent-written planner code, and figure-generation scripts for auditable policy adjustment in reinforcement learning control
Abstract
Evidence package for the manuscript "An Auditable Learning-Curve-Driven Automatic Policy Adjustment Framework for Engineering Reinforcement Learning Control Systems" (EAAI submission). The dataset contains the complete experiment ledger, the planner source code written by the diagnosis-driven agent, and the scripts that generate every figure in the paper. It supports three groups of reported results: 1. Radar primary domain (FluxPhased S7, a 2v2 radar-jammer adversarial system trained with MAPPO against frozen jammer pools, pool seeds 20260801-20260803): per-candidate resolved recipes, training metrics, canonical head-to-head evaluations, arm-level reports, and the literature cache holding both the diagnosis-matched and the diagnosis-irrelevant (shuffled-control) abstracts used by the content-ablation experiment. The ledger covers the random-search, BOHB, static-prompt, dynamic-retrieval, retrieval-shuffled, wide-box, and reflectively evolved-prompt arms, including the revision and completion rounds behind the retrained-winner comparison. 2. Code-as-Policy TD-MPC on 16 dm_control domains (four arms, three candidates each, seed 1301): all 192 planner sources (default CEM, diagnosis-driven agent, random parameter mutations, and diagnosis-free LLM generation), per-domain results, and the episode-level learning curves behind the learning-curve figure. 3. Seventeen-domain PPO generalization: agent-driven recipe loops and fixed-recipe baselines on 16 dm_control tasks, plus the HalfCheetah three-arm comparison. Model checkpoints (.pt) are intentionally excluded to keep the package light; every reported number can be regenerated from the recorded recipes, seeds, and metrics. Experimental environment: a single NVIDIA RTX 2060 (6 GB) at 7.72 s/iteration over 16 parallel environments; local Qwen3.5-4B language model (Ollama tag qwen35-4b-64k-full, Q4_K_M quantization). License: CC BY 4.0.
