The article is devoted to task planning in distributed information systems under conditions of uncertainty and incompleteness of metadata regarding the duration of operations, communication volumes, and resource availability. The subject of the study is the method of task planning in distributed information systems, taking into account the uncertainty of metadata. The purpose of the article is to experimentally investigate a method for planning tasks in distributed information systems with uncertain and incomplete metadata, which combines a hybrid graph model with a three-module artificial intelligence core and provides increased efficiency and robustness of planning while complying with SLO and RBAC constraints. Research objectives: to form a hybrid representation of workflows based on a combination of DAG structure and GERT model; to develop a planner architecture incorporating a GAT graph attention network, a PPO reinforcement learning agent, and a Bayesian Optimization module; to define an SLO-oriented reward function and constraint system, taking into account RBAC; to create a reproducible experimental test bed with controlled uncertainty injection; to perform comparative, ablation, and robustness analysis of the proposed method's productivity relative to the baseline FCFS and HEFT algorithms, as well as DAG-only and GERT-only variants. Research methods: hybrid graph modeling DAG+GERT, graph attention network GAT for forming a contextual state representation, PPO agent for making decisions on task placement, Bayesian optimization for tuning policy parameters, and statistical evaluation methods. Main results. The proposed method provides a significant reduction in the median task completion time (makespan) compared to HEFT, DAG-only, and GERT-only, an increase in the proportion of tasks performed within SLO, and a reduction in weighted average delay, demonstrates "tail compression" of the delay distribution, and maintains stability of results in the case of multiplicative noise of durations ±20 %, subject to distribution shifts and correlated parameter deviations. At the same time, planning overhead remains in the millisecond range and does not exceed acceptable limits for online scenarios.