The brittleness of transformer feature fusion: A comparative study of model robustness in engineering misinformation detection

Engineering decisions rely on the integrity of technical information, yet misinformation in design documents can cause catastrophic failures. We introduce the Engineering Misinformation Corpus (EMC), a large-scale multilingual benchmark curated from aerospace, energy, and civil infrastructure domains, and conduct a comparative study of feature-engineered, transformer-based, and hybrid fusion models for misinformation detection. Our results show that a feature-engineered XGBoost baseline and an end-to-end XLM-RoBERTa transformer achieve nearly identical performance on clean data. However, under adversarial perturbations, their behavior diverges. XGBoost degrades catastrophically, while XLM-RoBERTa remains the most robust. Naive feature fusion, defined as the direct concatenation of transformer embeddings with engineered features without dynamic learned arbitration, performs reasonably well, with only mild degradation. The gated fusion model, which introduces dynamic learned arbitration, avoids collapse but shows selective brittleness, with sharper drops under structural attacks. We conclude that while engineered features provide useful signals, both naive and gated fusion introduce vulnerabilities in specific adversarial settings, and transformers alone offer the most consistent robustness. This work establishes EMC as a reproducible benchmark, clarifies the trade-offs between the classical and transformer paradigms, and highlights open challenges for developing robust and trustworthy AI in engineering. All data, models, and code are openly released. • Introduces the Engineering Misinformation Corpus (EMC), a multilingual benchmark for technical integrity. • Compares a robust Transformer with a feature-engineered model and two hybrid fusion approaches. • Shows transformers outperform Engineered features by learning complex signals robustly. • Finds fusion models brittle; Gated arbitration fails under targeted semantic attacks. • Offers a reproducible benchmark with open-source data and code for trustworthy engineering AI.