Early-stage startups are central to innovation-driven economies, yet their failure rates remain persistently high, with more than half of new ventures not surviving their first three to five years. Accurately assessing the risk of young ventures is challenging because relevant signals are dispersed across heterogeneous sources, including entrepreneurial narratives, early financial indicators, and founders’ positions in entrepreneurial networks. Most existing models either focus on a single modality or treat multimodal features as independent inputs, overlooking whether cross-source discrepancies contain additional information and struggling with small, imbalanced datasets. This paper proposes a conflict-aware, generative-enhanced multimodal framework for predicting startup risk within 36 months after the initial institutional evaluation. The model jointly encodes business plan and interview text, structured financial features, and founder-centric social network information via transformer, multilayer perceptron, and heterogeneous graph neural network encoders. Cross-modal contrastive learning aligns the three modalities into a shared representation space, while a modal conflict attention module computes representation-level discrepancies between textual, financial, and network signals as candidate features for further inspection. To mitigate data scarcity and class imbalance, we further introduce a conditional GAN operating in the fused latent space to generate label- and conflict-conditioned synthetic representations that expose the classifier to both typical and strategically conflictual patterns. Experiments on a real-world dataset of 4,500 early-stage startups show that the proposed framework outperforms structured, text-feature, network-feature, and multimodal baselines. Compared to the best multimodal fusion baseline, our model improves AUROC from 0.82 to 0.87 and AUPRC from 0.53 to 0.62, with consistent gains in positive-class F\(_1\) and balanced accuracy. Ablation studies confirm the incremental contributions of contrastive alignment, modal conflict attention, and generative augmentation. A held-out construct-validity analysis further shows that the learned conflict representation tracks independently derived cross-domain discrepancy measures, while its non-monotonic outcome gradient supports using it as an inspection signal rather than a stand-alone diagnostic.