Expertium

Ok

So how would dynamic labels work? At epoch n we're doing optimization with some labels, then at epoch n+1 we're doing optimization with different labels? That seems like it would...

I tested an idea: making the hard_penalty act as a **multiplier** when it's >0 and as a **weight for a weighted average** when it's 0, it acts as a multiplier...

I don't understand how tanh would help here. And in my implementation w[15] is never supposed to be outside of that range, so what's the problem?

Ok, but why would that be a problem here, but not with all the other FSRS parameters, which are also clamped?

https://github.com/open-spaced-repetition/srs-benchmark/blob/8844f66c006a40ed71a409773468bd4a168b29c2/other.py#L938 

Also, just to clarify, [here](https://github.com/open-spaced-repetition/fsrs-optimizer/pull/173#issuecomment-2778505539) I didn't want to copy-paste the entire FSRS5ParameterClipper class just to show the difference in *one* line of code.

> When optimizing, the whole card is ignored. But while evaluating the parameters, only the reviews before the provided date are ignored. Wait, really? God, this is even more of...

Wait, I thought that for scheduling, no reviews are ignored, and cards just use parameters that were generated by the optimizer while ignoring reviews? So the reviews are ignored for...

Ok, so there are only three people in the world who understand how "Ignore reviews before" works: @L-M-Sherlock, @user1823, and @Luc-Mcgrady. Unfortunately, LMSherlock's English isn't good enough to write the...