Scientific ML: Scientific image forgery detection

The agent classifies biomedical research images as authentic or manipulated and segments every copy-moved region as a separate instance, training offline and submitting through a sealed leaderboard gateway.

Image manipulation distorts the scientific record, and instance-level localization gives an integrity reviewer far more to work with than a bare "manipulated" flag. Biomedical images raise the difficulty: the textures are repetitive by nature, so distinguishing a cloned region from an honestly similar one is the whole problem.

Scoring uses an F1 that combines authenticity decisions with instance-mask matching. Scored submissions are quota-limited and return public scores only; the reward is the floor-normalized private score of the best-public submission.

This task was built independently from the same source problem as copymove, with a different split and floor. The two scores are not comparable.

The strongest run committed early to a single self-correlation model and improved it in place: higher-resolution inputs, then a confidence gate on detections, then continued training. Test-time augmentation came at the close. It rejected its ensemble attempts because the members were too correlated or simply weaker. It made the fewest scored submissions and finished earliest.

The weakest run built a verifier-style pipeline. Its last logged ideas were all rejected: an over-conservative ensemble, an evidence-gated accept rule, extra hypotheses. Its final hour of submissions cycled near-identical ensembles of the same verifier variants for less than a point of public gain.

The middle run spent nearly three hours building a dense-matching pipeline with a refiner stage while barely submitting. It recalibrated its decision threshold in one coarse pass, then burned two dozen scored submissions on fine threshold and overlap tweaks. Those public points mostly failed to carry to the private split, leaving it the widest public-private gap on the task.