PCaBiop (slide-level)

1,000 prostate biopsy slides sourced from PANDA, labelled benign or cancer, contributed by two data providers (Karolinska Institutet and Radboud UMC). Unlike the three tile cohorts, the evaluation unit is a whole slide: each of the 5 encoders is a whole-slide foundation model producing one embedding per slide, and every number below is computed over those slide embeddings.

This is a different panel. The slide roster shares no encoder ranking with the 26-model tile panel, so this cohort takes no part in the cross-cohort aggregate and its ranks — the table below orders these five encoders against each other, nothing more. There is also no natural-image control at slide level, so no row carries the † mark.

PCaBiop, sorted by median CRoMa. Columns are explained under Reading the columns.

Model	bio bacc	conf bacc	RI	MaRI	Δ	CRoMa	F(0)	LTM₁₀	support
PRISM2	0.975	0.995	0.434	0.367	-0.067	0.27	0.187	-0.24	20.9%
PRISM	0.968	0.992	0.281	0.195	-0.087	0.26	0.288	-0.39	10.1%
MOOZY	0.971	1.000	0.236	0.181	-0.054	-0.02	0.535	-0.41	24.3%
TITAN	0.915	1.000	0.015	0.001	-0.013	-0.30	0.895	-0.60	49.1%
Prov-GigaPath	0.720	0.995	0.020	0.006	-0.014	-0.41	0.990	-0.59	97.7%

The confounder is unusually dominant here: every encoder’s confounder k-NN accuracy is at or near 1.0 — the data provider is essentially perfectly decodable from every slide embedding — and 3 of the five encoders hold a negative median margin. Support spans 10% to 98%, the widest spread of any published cohort, so the count-based indices rest on very different amounts of evidence per row: Prov-GigaPath’s low RI is measured on nearly every slide, while PRISM’s rests on a tenth of them.

The two rankings, on this cohort alone:

Median CRoMa against tail severity LTM₁₀. Better is up and to the right; ringed points are undominated on both axes. Hover or tab to any point to name it with its two values. With five encoders the panel is sparse by construction — read it as a picture of the table, not as a frontier claim of tile-panel strength.

The distribution explorer

The same explorer as the tile panel’s, restricted to the slide roster. Click a row to move the detail, drag across the large histogram to count the slides in any range, and pick a second encoder under Compare with to overlay its shape.

The shape is again where the medians stop telling the story. PRISM2 and PRISM sit 0.014 apart on median CRoMa — indistinguishable on that column — while PRISM carries 1.5× the confounder-dominant mass and 1.7× the tail severity. Overlay the two above to see it.

The operating point

This cohort is reported at k*, not the tile panel’s shared median-k: each encoder is evaluated at its own kNN-optimal k (here spanning k = 3–13; the per-encoder values are recorded in results/PROVENANCE.json). With only five encoders, a shared median is dominated by panel composition — adding a single encoder to the panel moved the would-be shared k from 3 to 9 — so pinning one k would make every number hostage to who else happens to be on the roster. The trade is stated rather than hidden: RI and MaRI are protocol-dependent, so those two columns are not measured at one shared operating point here, while CRoMa, F(0) and LTM₁₀ are k-free and unaffected.