The weak point in driver identification is rarely the sensor. It is the assumption that every new driver entering a fleet requires retraining a classifier. TEMPEST sidesteps the problem at the root: it projects each driving window into a normalized angular space, where identity becomes a matter of distance between vectors, not re-labeling.

The architectural choice is clear. A Temporal Convolutional Network learns compact 96-dimensional embeddings from 60-second multimodal windows, using an additive angular margin loss, ArcFace. The result is not a classifier to update, but a gallery of reference vectors. The model, with 720K parameters and a 2.80 MB footprint, converges in 50 epochs. This is the technical premise that enables dynamic enrollment: adding or removing a driver requires neither retraining nor classifier refitting.

On 45 drivers, TEMPEST reaches 91.71% Rank-1 accuracy. The gap from the best classical model is 17.9 percentage points; from the strongest triplet-loss baseline it is 58.4 points. But the number that best captures the problem is the degradation as the subject pool grows: moving from 10 to 45 drivers, TEMPEST loses 4.3 percentage points, while supervised and unsupervised triplet-loss baselines lose 22 and 32.5. On the public Kia Soul dataset, the advantage is 7.3 points within-session and 14.3 points cross-session over the best classical model.

The difference between triplet-loss and ArcFace is not an academic nuance. Triplet-loss learns relative relationships between examples and tends to overfit session-specific patterns under rigorous temporal evaluation. Additive angular margin instead enforces global class separation in a normalized space. In practice, each identity occupies a stable angular region, not a snapshot of a specific session. That difference explains the model's retention as the number of subjects grows.

For fleet operators, the implication goes beyond performance. Dynamic enrollment changes the cost of identity management. A new driver becomes a vector added to a local database, not a label injected into a training pipeline. This shifts the deployment center of gravity toward the edge: a 2.80 MB model can live inside the vehicle or on a fleet server without sending driving data to a central cloud. The losers in this scenario are approaches that assume periodic centralized retraining and vendors that monetize that cycle. Those evaluating on-premise deployment can find analytical frameworks on AI-RADAR's /llm-onpremise to compare trade-offs, but the core point here is that data sovereignty stops being a luxury if enrollment and inference stay local.

The evaluation on 45 drivers does not justify projecting the result onto fleets of thousands of vehicles, and the work does not claim to do so. The strength is the protocol: rigorous temporal evaluation, cross-session comparison on a public dataset, measured degradation as the subject pool expands. It is a reproducible foundation for behavioral biometric identification. Structurally, TEMPEST signals that driver identification is converging with the metric learning already widely used in face recognition. There is no longer a need to train a classifier for each fleet; the task is managing a gallery of embeddings in a well-separated angular space.

This does not remove open questions about robustness to noisy multimodal signals or much larger driver pools. But the fact that a compact, reproducible model limits degradation under strict temporal evaluation is a concrete signal: the bottleneck in biometric identification is no longer compute power, but the quality of the latent space and the discipline used to evaluate it.