Asset specificity critically determines governance structures under incomplete contracting. While transaction cost economics prescribes hierarchical governance to protect specific investments, this view overlooks their potential to drive technological change. This paper models how asset specificity evolves and facilitates the emergence of General Purpose Technologies (GPTs). We show that governance focused on static efficiency can suppress innovation, creating a trade-off between appropriation protection and innovation. Under uncertainty, firms choosing market governance despite high asset specificity enable systemic spillovers and technological generalization. The model integrates transaction cost economics, capabilities theory, and GPT literature, reinterpreting relational contracts as governance mechanisms for dynamic efficiency.

This paper proposed an outlier removal method for magnetoencephalography (MEG) data during auditory imagery tasks to improve classification accuracy. While previous studies on imagined speech brain-computer interfaces (BCIs) mainly relied on electroencephalography (EEG), MEG offered higher spatial resolution but suffered from smaller data sizes, making it more sensitive to abnormal trials. To address outlier removal, we first pretrained a feature extractor on MEG data that had been augmented based on current source estimation. Second, we applied the trained feature extractor to obtain feature representations of the MEG samples. Third, these features were projected into a low-dimensional space using t-SNE. Outliers were then identified based on their deviation from the feature distribution and were removed before classifier retraining. Experimental results demonstrate that the proposed method improved classification accuracy, particularly when applied to both training and validation data. Furthermore, training the feature extractor on data augmented based on current source estimation likely led to clearer class boundaries, which further contributed to the performance improvement. These findings suggest that data augmentation based on current source estimation and t-SNE-based outlier detection enhance the robustness of MEG-based imagined speech decoding. © 2025 IEEE.