Ordered roughly by expected value for the research question.

1. Uncertainty in the numbers themselves. Repeated grouped splits with paired ablations are implemented (scripts/evaluate_multisplit.py); show them on the website's Experiments page as well, and extend them to the collection-disjoint experiments.
2. More institutions. The Art Institute of Chicago (its API was blocked for programmatic clients during development; its CC0 data dump is an alternative), the Rijksmuseum, the Smithsonian Open Access, Europeana providers with explicit CC0/PDM rights. More institutions make collection-disjoint evaluation far more informative.
3. Better out-of-distribution evaluation. A held-out set of non-paintings (sculpture, photographs, prints) and of traditions absent from training, to measure the abstention and out-of-scope rules directly.
4. Representation comparisons. Fine-tuned DINOv2 (linear-probe vs. partial fine-tuning with grouped cross-validation), other self-supervised encoders, multi-crop or tiled embeddings for extreme aspect ratios (handscrolls).
5. Ordinal and range-aware period models. Ordinal regression or interval-censored losses that use date ranges directly instead of discarding works that span several bins.
6. Hierarchical modelling. Joint region–tradition models; conformal prediction sets ("one of these two regions") instead of a single abstention threshold.
7. Explanations that are actually computed. Patch-level similarity maps between an upload and its nearest neighbours (DINOv2 patch tokens), shown as evidence of where images are similar — with the caveat that similarity is not reasoning.
8. Other media. Prints, textiles, ceramics, sculpture — each with its own photographic conventions and evaluation.
9. Curator-in-the-loop review. A reviewed sample of region mappings and failure cases (see annotation_protocol.md) to estimate label noise.