Enrichment by variance collapse

the finding

More Events from a Narrower Distribution

Classifier guidance conditions the emulator’s sampler on the events of interest, so rare outcomes appear in the sample far more often than they would unguided. The reported enrichment factors of 4 to 90 are real, and they are the method’s pitch: tail statistics at a fraction of the compute.

The enriched samples do not come from the same tail. Pooled across events, the fitted generalised Pareto shape parameter of guided samples is −2.005, compared with −0.265 for unguided ones: a sharply truncated tail in place of a mildly bounded one. In other words, the guidance concentrates probability mass near its target rather than extending the model’s reach past it. That collapse of scale is what makes the extra events cheap.

The method’s own reliability diagnostic agrees: the estimator that turns guided samples back into event probabilities fails its own k gate on these runs, so the enriched sample does not yield a trustworthy exceedance estimate.