Trimming-based consensus rules such as weighted mean-subsequence-reduced (W-MSR) consensus protect multi-agent networks against a bounded number of faulty or malicious agents, but most of their guarantees assume synchronous, delay-free communication. Real networks deliver delayed, reordered, and replayed messages, and timestamps cannot be trusted when adversaries can forge them. For resilient consensus with bounded delays, prior work shows that (F+1, F+1)-robustness is necessary and (2F+1)-robustness is sufficient. To our knowledge, whether the gap between the two conditions is real has remained open. We study a timestamp-free W-MSR rule in which every normal agent trims received values relative to its own current state while neighbor values arrive with arbitrary time-varying delays bounded by τ. First, we prove a safety property that holds for every adversary and every delay pattern: the hull of normal states over the most recent τ+1 steps never expands. Second, we prove that (2F+1)-robustness suffices for exponential consensus, with a contraction horizon of |N|(τ+1)+τ steps that is linear in τ+1. Third, we show that (F+1, F+1)-robustness is not sufficient once delays can be scheduled adversarially: with τ=1, a broadcasting adversary can present different recent values to different receivers and block consensus on a family of graphs that, by exhaustive enumeration, is (F+1, F+1)-robust for F ≤ 2 and 2F-robust for F ≤ 3. We also specify a reproducible simulation protocol built on certified robust graphs, three delay schedules, and four adversary models including stale replay. This preprint presents the model, the analysis, and the evaluation protocol. It does not report empirical outcomes.

Cooperative multi-agent reinforcement learning (MARL) relies on communication to coordinate under partial observability. Real channels lose, delay, and corrupt messages, and a compromised agent can send misleading ones. In hierarchical MARL, these failures propagate into task-allocation errors and unsafe low-level actions. Prior work has studied message filtering, communication-constrained learning, and hierarchical recovery under message loss largely in isolation. This paper proposes RiCoR, a risk-calibrated communication and recovery framework for hierarchical MARL. RiCoR separates observable channel conditions from message credibility and represents uncertainty in its reliability estimates with an evidential Beta posterior. It gates each message by a lower confidence bound whose threshold rises with the downstream decision risk, measured by a conditional value-at-risk (CVaR) cost critic. A learned mode selector switches among normal coordination, selective reliance, active verification, and local fallback, with hysteresis and explicit switching and verification costs. A one-step safety shield filters low-level actions, and its guarantee is stated under explicit assumptions. The paper also specifies an evaluation protocol for a hierarchical wildfire-suppression benchmark with one coordinator and eight regional agents. The protocol covers isolated, compound, and held-out faults, four baselines, and calibration, recovery, ablation, scalability, and overhead analyses. The paper reports no empirical outcomes.