A fair skeptical question: did you just look at the known labels and tune a rule until it matched them? This page explains one specific, structural reason why that's unlikely here — and why it's different from how an AI cancer predictor works.
This is a general problem in all quantitative science, not specific to AI: if a classification threshold is adjusted after looking at which samples are cancer and which aren't, it's expected to separate that same data well — that alone doesn't prove the rule captures anything real. This is why the source of a "no exceptions" claim matters as much as the claim itself.
A supervised AI cancer classifier is trained per task: it needs labeled examples for a specific species, a specific cancer type, often a specific tissue site, and typically needs retraining or fine-tuning to generalize to a new one. This dependency is a direct consequence of how such models are built — they are fitted to the statistical patterns of one labeled category at a time.
A key cut specifically to fit one lock usually won't open an unrelated lock. A key that opens many unrelated locks, without being re-cut for each one, is more likely to be exploiting some real, shared mechanism in how those locks are built — not a coincidence.
If the Tail Runaway threshold had been tuned to match the labels in one dataset, there would be no particular reason for that same, unmodified rule to also separate cancer from non-cancer in a completely different species, using a completely different tissue-collection pipeline, without any species-specific or cancer-type-specific adjustment. The fact that one fixed rule holds up across human, rat, and dog data, and across 33 independently collected TCGA cancer types, is evidence against the rule being an artifact fitted to any single dataset's idiosyncrasies.