A One-Year Value Target Read Three Quarters of the Decline After 28, Every Season

The last valuation 180 to 365 days after a season had a median horizon of 225 days and read growth as -2.9% where a year gives -7.5%. Dropped rows explain 0.3 points. The revaluation calendar does the rest, and it compresses every season alike, so no out-of-time check on the same target can see it.

Until December, Last Season's Form Prices the Market Better Than This Season's

A season replayed on four dates. Last season's form alone fits the market at R² 0.757 after five games against 0.610 for this season's, still leads in December, and is the furthest of three variants from the model's verdict on the full season.

Shift the Window Grid 100 Pixels and 64% of a ViT's Detections Change

A ViT mitosis detector and a CNN of the same size, same head and same training, run over whole tissue regions as overlapping windows. Moving the window grid by 100 px changes which side of the threshold 64% of the ViT's detections fall on, against 9% for the CNN.

Colour Is Worth 0.24 AP on the Training Tumours and 0.10 on the Unseen One

Four mitosis classifiers trained in colour, evaluated on greyscale copies of the same crops. On the tumours they trained on, removing colour costs 0.24 average precision. On a held-out tumour it costs 0.10, and the domain gap in ranking changes sign.

Ten Hours of LeJEPA Buy 0.04 Over Six Colour Numbers

A ViT-Tiny pretrained with LeJEPA on 174,110 histology tiles shows no collapse, an isotropic embedding and a loss that falls for all 100 epochs. Its class token predicts tumour type at 0.755. Six colour statistics of the same window score 0.717, and the untrained encoder 0.714.