We present a benchmark-driven methodology to study how historically-inspired synthetic degradations affect the performance of video restoration models. Using a modular pipeline, we generate artificial artefacts on clean REDS [6] sequences and evaluate their impact on restoration quality. Our goal is to improve restoration robustness by selecting and composing degradation types that better reflect real archival damage. We compare RRTN [4] against variants trained on different degradation presets and provide empirical guidance for effective model training when paired data are scarce and contributes to the discussion on data efficiency in small-data scenarios.