With yet more help from Gabriele, we've got a freshly trained LSC cleaning, using the same training data as alog 70023.

The 'secret tricks' were to actually sample the 16384 Hz GDS filter at 16384 Hz (rather than my mistakenly using 2048 Hz), and then to train without using the ASC modulations at first to get good starting estimates for filters, then fine tune the training with the modulations added back in. 

Attached is the offline subtraction that we should be able to achieve.  Note that this LSC training will be in competition around 120 Hz with the Jitter training from alog 70023, so I'll likely have to implement one of them, collect a few mins of data, and retrain on that data.

I'll try this out during our commissioning window this afternoon.