Test the floor-effect explanation for mechanism-dependent ageing

Test whether the slower apparent ageing of the TypeNet-style LSTM embedding model results from a floor effect caused by poor representational quality, and whether improved training would cause its ageing slope to steepen while the more accurate TypeFormer-style Transformer exhibits greater genuine degradation.

Background

The paper finds that the TypeNet-style LSTM has high baseline error but a robustly shallower ageing slope than the scaled-Manhattan matcher, whereas the TypeFormer-style Transformer has lower baseline error and tentative evidence of faster ageing under the logistic specification.

The authors propose a floor-effect account: if much of M3’s error is due to limited representational quality rather than temporal drift, it may have less remaining error headroom to increase as the enrolment-to-query gap grows. They explicitly characterize this explanation as unverified and identify it as a target for follow-up work.

References

We lack a confirmed mechanistic explanation for this, but we can offer one testable hypothesis, which remains explicitly unverified.

— Template Ageing and Longitudinal Verification in Fixed-Text Keystroke Dynamics: A Subject-Disjoint Study Across Eight Weeks  (2609.29851 - Parkinson et al., 24 Sep 2026) in Section 5.1, “Mechanism-Dependent Ageing Rates,” and Section 5.2, “Limitations”

The ageing result, finally, is measured but not explained. We show that the ensemble ages at its members' rate and that the small student ages more slowly, without knowing what drives either.

— Unknown-Traffic Detection, Calibration and Shortcut Reliance in Distilled Encrypted-Traffic Classifiers over One Year  (2609.31141 - Abbasi, 25 Sep 2026) in Section 7, “Conclusion and Future Work”