HARD PROBLEMS · 2026 · CANDIDATE FALSIFICATION · EXPERT-CONFIRMED JOINS 0
The system ranked possible clay-tablet fragment joins. It did not establish a new join. Known pairs served as positive controls, and the work produced comparison candidates and an image bundle. The model’s discrimination was not decisive, and no specialist confirmed a new physical connection.
Candidate identifiers and collection images are withheld. Their reuse terms have not been cleared, and an unreported candidate should not be made to look like a discovery through publication design.
Positive control
known joins
Candidate ranking
internal output
Model decision
inconclusive
Expert-confirmed joins
0
The method first had to recover a join already known
A known pair was placed in the evaluation as a positive control. If the method could not rank that pair convincingly, there would be little reason to trust its ordering of unseen fragments. The control was a minimum test of what the model was detecting.
The next stage ranked candidates using shape and edge comparisons and assembled views for human inspection. A list can change the order in which a specialist looks. It cannot establish that two damaged objects physically join.
A falsification record mattered more than a high rank
Two fragments may look similar yet disagree in thickness, fracture surface, writing direction or archaeological context. Without recording those failures candidate by candidate, model ranks can turn too easily into human confidence.
The current evaluation split and performance evidence are not strong enough for a public result. No new join has specialist confirmation. Claims of a discovered join, successful restoration or high accuracy are excluded.
A candidate rank can reorder a specialist’s queue. It cannot decide that two fragments once formed one object.
Image rights come before a defensible restart
Follow-up work needs images licensed for the intended reuse, a clean separation between development and evaluation material, and candidate-level falsification notes. A domain specialist would then need to review the primary objects and archaeological context.
Online visibility is not permission to reproduce a collection image in a portfolio. Until a specific asset’s terms are verified, this lineage stays text-only and the candidate list remains private.
Project record
- Period: 2026
- Activity: positive controls from known joins, candidate ranking, comparison views and falsification design
- Internal evidence: candidate ranks, comparison-image bundle and model output
- External validation: none
- Final outcome: zero expert-confirmed new joins
- Not established: general model performance, physical fit, image-reuse rights or specialist judgment