Meta’s Stella companion application ships with a fully assembled, dormant facial recognition pipeline, according to technical analysis of version 273.0.0.21 for Android. While the feature remains inactive for standard accounts, the complete computational and storage infrastructure required for on-device face matching is present and functional. Dated June 4, 2026, the findings detail an end-to-end biometric stack integrated into Meta’s smart glasses software. The pipeline comprises three ExecuTorch models totaling approximately 100 megabytes. An SCRFD detector isolates faces within captured imagery, a KPS aligner crops and standardizes them, and an SFace model generates a 2048-dimensional biometric embedding. These embeddings are processed through a local SQLite database utilizing a cosine-similarity vector index. The system supports both successful matches and unidentified faces. Recognized individuals trigger a high-importance Android notification displaying a predefined title and the matched contact’s name. Unrecognized faces are staged to a local directory as paired image and embedding files, formatted identically to the live recognition index, indicating a design for future retroactive labeling. Despite the complete backend architecture, active deployment remains unverified. On a stock, unenrolled account, the user interface does not expose facial recognition controls. A hardcoded Connections widget exists within the application package but remains hidden under standard viewing conditions. Furthermore, the Android notification includes a deep link to a profile screen that is currently absent from the navigation graph, causing the app to revert to its default view upon interaction. The underlying person-profile database resides within Meta’s cross-device RLDrive sync framework, yet investigators observed no evidence of server-side identity data being pushed to this namespace during testing. The discovery underscores a significant engineering investment in biometric identification infrastructure. The detection, alignment, embedding, vector indexing, storage routing, and notification surfaces are mutually consistent and operationally coherent. This configuration suggests that Meta has developed a complete on-device recognition capability intended for future rollout or specific user enrollments. The architecture does not confirm active surveillance of bystanders, nor does it indicate that biometric data is currently being harvested or pushed in production environments. Instead, it demonstrates that the foundational machinery for smart glasses-based facial identification has been built, integrated, and shipped, awaiting administrative enablement or user enrollment. As reported alongside WIRED, the findings highlight the gap between deployed technical capability and consumer-facing activation, leaving questions regarding Meta’s timeline and privacy safeguards for on-device biometric matching unresolved.