Built to say “I don't know”.

We publish our failure modes on purpose. A model that hides them is more dangerous than one that names them. These come straight from our model card.

• External performance is below the internal number. Macro-F1 is 0.940 on the held-out test set (n=2,700, GDC, 25 sites), but on independent external tumors the honest figure is lower: a 7-site top-1 of 0.906 on 381 cBioPortal samples. We report the external number separately rather than blend it in to look better.
• Cross-platform inputs are harder. Calibration and abstention guarantees hold in-distribution and degrade on data from a different sequencing platform or pipeline. The model abstains on inputs it does not recognize rather than guess.
• Some sites are genuinely weak. Rare sites with little training data (Thymus, Pleura and Mediastinum, Esophagus, Skin) are less reliable, and a few sites are overconfident on real patients (for example Soft Tissue and Esophagus, where stated confidence runs well above measured accuracy).
• Mesothelioma is a known miss, and it fails silently. An earlier pleural-cancer result turned out to be a batch artifact: on real TCGA mesothelioma the model currently has 0% recall. Worse, it does not abstain on these cases. It confidently mislabels real mesothelioma as Thymus (at roughly 0.76 confidence), so the error looks like a normal, sure-footed call rather than an uncertain one. We do not claim mesothelioma detection, and a confident Thymus call should not be trusted to rule mesothelioma out.
• Training data skews toward certain populations. The training cohorts (GDC, predominantly TCGA-derived) are not characterized for race or ethnicity, ancestry, age, sex, or geographic balance, and are known in the literature to skew toward certain populations. Performance across underrepresented groups is unmeasured and may be materially worse.
• No subgroup fairness audit yet. We have not completed a fairness audit across sex or genetic ancestry, so we make no subgroup performance claims.
• Accuracy depends heavily on the source cohort. On raw external cohorts from different sequencing centres (before our harmonization step), measured per-source accuracy ranged from about 6% to 100% across eight independent cBioPortal studies. This is why the model harmonizes inputs and is designed to abstain on profiles it does not recognize rather than guess, though that safeguard is not absolute and a confident call on an unfamiliar cohort is not a guarantee of accuracy. Do not assume uniform accuracy across cohorts or platforms.
• Everything downstream is a hypothesis. Subtype, druggable-target, and drug-response outputs are research hypotheses for expert review, not conclusions, and never treatment guidance.