Could decorrelation miss new physics if that physics couples strongly to the protected variables you chose to ignore?

Yes—and this is exactly the core tension that decorrelation methods like those in [AG-2024.09-1051] are designed to *acknowledge* rather than solve.

Here's the worry in plain terms: if you decorrelate your classifier from some "protected variable" (say, the mass of a particle) to avoid being fooled by background model errors, you're implicitly assuming that any new physics you're hunting for *doesn't* couple to that mass variable. But if it does couple strongly, you've just thrown away your signal.

[AG-2024.09-1051] addresses this by using decorrelation only for a preliminary enrichment step, then performing the actual signal detection on the mass distribution itself—rather than relying on the decorrelated classifier alone. The idea is: use the classifier to collect more signal-rich events, but then do your inference in the variable you decorrelated from. This way, if new physics *is* coupled to that variable, it would show up as an excess in the mass spectrum of the enriched sample.

However, the archive materials here don't directly explore what happens if new physics couples to variables you never even *considered* protecting. The papers on new physics searches [AG-2025.05-1089], [AG-2025.01-1105] use quantum information measures to look for anomalies, but they aren't framed as a response to the decorrelation risk.

The deeper point: decorrelation is a tool for *background* robustness, not signal discovery. It protects you against your background model being wrong—but you still need to be thoughtful about which variables matter for the signal you're seeking.

The preprints don't directly address how to identify which variables might couple to hidden new physics in advance.