Resolve the provenance and training budget of the evaluated FARE checkpoint

Determine which upstream FARE ViT-B/32 checkpoint produced the locally evaluated file and thereby establish its correct adversarial-training budget for interpreting the regime mismatch between training and evaluation.

Background

The experiments use a locally stored FARE checkpoint whose hash matches the frozen experimental manifest. However, the artifact-repository README labels it as trained at an epsilon of 4/255, while the RobustVLM release lists multiple FARE ViT-B/32 variants, including 1/255 and 4/255 versions. Because the evaluation budgets are compared with the checkpoint's training budget, identifying the exact upstream checkpoint is necessary to support the paper's interpretation of the training–evaluation regime mismatch.

References

Our FARE checkpoint is labeled $\epsilon_{\mathrm{train}{=}4/255$ based on the artifact-repository README, but the RobustVLM release lists multiple FARE variants for ViT-B/32 ($\epsilon_{\mathrm{train}{=}1$ and $4/255$), and we have not independently verified which upstream checkpoint produced our local file. The checkpoint's SHA-256 hash matches the frozen manifest, so the experiments are internally consistent, but the training-budget attribution and the regime-mismatch interpretation depend on resolving this provenance question.

— Does Target Alignment Mean Target Recovery? An Evidence-Ladder Study of Adversarial Claims on Contrastive Encoders  (2610.11938 - Yang et al., 8 Oct 2026) in Section 6, Discussion—Checkpoint provenance