Abstract: To address the limited regional adaptability of fixed-parameter filtering in seismic random noise attenuation, the dependence of supervised learning on clean labels, and the lack of local decision interpretability and processing-intensity control in self-supervised static mapping, this study proposes a region-guarded blind-reward sample-level reinforcement learning method, termed BRGF-PRL. The proposed method formulates denoising as a finite-horizon sample-level sequential decision-making process. Blind priors, including event confidence, noise confidence, weak event support, and neighborhood consistency, are constructed from noisy observations. A region-guard map, namely a weighted mask designed to protect valid reflection structures from excessive attenuation, is introduced to regulate the update intensity at different spatial locations. Meanwhile, a blind reward that integrates noise suppression and structural preservation constraints is designed to guide policy optimization. Four sets of controlled synthetic experiments demonstrate that BRGF-PRL achieves the highest average SNR improvement and average SSIM among the compared methods without clean labels, reaching 12.54 dB and 0.986, respectively. Ablation experiments further verify the major contributions of the region-guard mechanism and adaptive fusion strategy. Experiments on real data windows further show that the proposed method produces less leakage of effective reflections in terms of profile morphology, residual characteristics, and spectral behavior, achieving a stable balance between random noise suppression and reflection-structure preservation.