<p>The rapid advancement of the mobile internet and the internet of things has led to the emergence of many latency-sensitive tasks at the terminal side. The traditional three-layer MEC architecture of "terminal-edge server-cloud" has limitations in processing speed, deployment cost and coverage. To address this, we propose a task offloading mechanism that uses trusted energy consumption assessment for edge collaborative computing with inter-terminal collaboration. We first construct a two-layer "edge-cloud" architecture and create a MILP formal model for task-node-link interactions. This highlights challenges in optimizing both latency and energy consumption. Next, we design a node evaluation mechanism that combines self-assessment, peer assessment, and trusted endorsement. The Analytic Hierarchy Process (AHP) and Dynamic Grey Relational Analysis (DGRA) are utilized to calculate indicator weights and similarity scores. These outputs are synthesized into a trust-aware energy-efficiency score that strikes a balance between interpretability and discrimination. The score is applied in a Markov Decision Process (MDP) using a feasibility mask, which enforces strict constraints on unreliable or costly nodes at the action space level. We also present the evaluation-masked and attention-enhanced multi-agent deep reinforcement learning (EA-MADRL) framework. It utilizes a multi-agent division of labor and a dual-layer multi-head self-attention mechanism to capture temporal fluctuations and key task-node relationships. In simulations against baselines such as MADDPG, MASAC, and MAPPO, alongside several ablation variants, our method reduces average latency and boosts returns under medium to high loads or low cloud parallelism. When resources are abundant, it maintains steady energy consumption reduction without causing extra latency. The EA-MADRL parameter determination optimization process converges faster and is more robust.</p>