<p>Social recommendation commonly uses observed social relations to alleviate sparse user-item interactions. However, observed social relations do not necessarily encode preference information useful for a given recommendation task. Their utility depends on the alignment between behavioral patterns and social relations, as well as user-specific conditions. Most existing methods focus on propagating, fusing, or denoising social information, whereas the task relevance of social relations and user-specific control over social-signal injection remain less explicitly modeled. To address these issues, this paper proposes Conditional Adaptive Reliability Estimation (CARE) for social recommendation. CARE uses behavioral representations as task-relevant references to estimate the reliability of social relations and adaptively regulates the strength of social information injection according to user-level behavioral and social conditions. As an auxiliary stabilization mechanism, CARE assigns stronger conditional consistency regularization under high-confidence social conditions and reduces its strength when social reliability is low, thereby helping avoid overly strong constraints under low-reliability conditions. Experiments on three public datasets show that CARE generally achieves competitive performance under different social data conditions. CARE obtains relatively larger gains when social relations are more consistent with behavioral preferences, while remaining comparable to or slightly better than strong behavioral baselines under weaker social-behavior consistency or higher social noise. Ablation experiments and user-level analyses further suggest that its gains are mainly associated with task-relevant reliability estimation and user-specific regulation of social injection, rather than uniform enhancement of social information. These results support modeling social information as a conditional auxiliary signal rather than a fixed, inherently effective source.</p>