| <p>Ordered roughly by expected value for the research question.</p> | |
| <ol> | |
| <li><strong>Uncertainty in the numbers themselves.</strong> Repeated grouped splits with paired ablations are implemented | |
| (<code>scripts/evaluate_multisplit.py</code>); show them on the website's Experiments page as well, and extend them to the | |
| collection-disjoint experiments.</li> | |
| <li><strong>More institutions.</strong> 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.</li> | |
| <li><strong>Better out-of-distribution evaluation.</strong> 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.</li> | |
| <li><strong>Representation comparisons.</strong> 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).</li> | |
| <li><strong>Ordinal and range-aware period models.</strong> Ordinal regression or interval-censored losses that use date ranges directly | |
| instead of discarding works that span several bins.</li> | |
| <li><strong>Hierarchical modelling.</strong> Joint region–tradition models; conformal prediction sets ("one of these two regions") | |
| instead of a single abstention threshold.</li> | |
| <li><strong>Explanations that are actually computed.</strong> Patch-level similarity maps between an upload and its nearest neighbours | |
| (DINOv2 patch tokens), shown as evidence of <em>where</em> images are similar — with the caveat that similarity is not | |
| reasoning.</li> | |
| <li><strong>Other media.</strong> Prints, textiles, ceramics, sculpture — each with its own photographic conventions and evaluation.</li> | |
| <li><strong>Curator-in-the-loop review.</strong> A reviewed sample of region mappings and failure cases (see annotation_protocol.md) | |
| to estimate label noise.</li> | |
| </ol> |