Far more details are found in the vignette. At any time point tau, the
scaled Brier score is essentially the R-squared statistic where y =
the 0/1 variable "event at or before tau", yhat is the probability of an
event by tau, as predicted by the model, and the ybar is the predicted
probablity without covariate, normally from a Kaplan-Meier.
If \(R^2= 1- \sum(y- \hat y)^2/\sum (y- \mu)^2\), the Brier score is formally only
the numerator of the second term. The rescaled value is much more
useful, however.

Many, perhaps even most algorithms do not properly deal with a tied
censoring time/event time pair. The `tied`

option is present
mostly verify that we get the same answer, when we make the same
mistake. The numerical size of the inaccuracy is very small; just large
enough to generate concern that this function is incorrect.

A sensible argument can be made that the NULL model should be a
`coxph`

call with no covariates, rather than the Kaplan-Meier;
but it turns out that the effect is very slight.
This is allowed by the `efron`

argument.