Results
Machine-generated results and human corrections remain separate.
Case study no. 2
LoraCrafft 1.0.48 · Assessment dated 17 October 2026 · inscription and cartouche.
| Measure | Value |
|---|---|
| Automatic regions | 24 |
| Reference occurrences | 18 |
| Matching codes on the proposed alignment | 13 |
| Code disagreements | 5 |
| Regions without a separate reference token | 6 |
These counts are not an overall recognition accuracy. Human ground-truth boxes are missing and several differences concern sign scope or grouping.
Recognition: G1/G5, N24E/N16 (×2), M156/M23, I11A/I10.
Structure: the two unions below F36 are each retained as one unit; the scope of M23 and several additional signs remains unresolved. Whole-inscription MdC is not validated.
Data and human review
The human reference and draft annotations were prepared separately, without rewriting predictions or adding Dice references. These pages distribute no corpus, private document or source image.