Hello, I am developing a Deep Learning Survival Model (similar to DeepSurv). In my preprocessing step, I have used the quantile transformation on my covariates, and I have considered two choices for my target distribution as follows:

According to what I know, both approaches start by transforming the value of features into their empirical quantiles, so they both preserve the original rank ordering within each feature. The only difference is the second step: the uniform option keeps the quantiles on a bounded scale

In my experiments, using the uniform output consistently gives a higher C‑index than the normal output for the same architecture, loss, and hyperparameters. My current hypothesis is that, for this particular deep network and optimizer, having all features in a tight, bounded range makes optimization easier, whereas the normal mapping introduces heavier tails/unbounded values that make training a bit less stable, even though both transformations preserve ranking.

My questions:

1. Are there known reasons why a bounded uniform quantile transform might empirically give better C‑index than a normal quantile transform in deep survival models?
2. Is it reasonable to attribute this mainly to optimization dynamics (scale/tails of inputs) rather than to loss of information or changes in ranking?
3. Are there recommended diagnostics (e.g. checking calibration metrics like Brier score, trying batch/layer normalization, etc.) to systematically compare these two preprocessing choices?