"""CAISc Paper — Transformer Training Pipeline.

Trains encoder-only transformers on HH circuit simulation data to predict
population-level summary statistics.  Two models:

  Model A  – plain HH (ACh = 0 only)
  Model B  – HH + ACh modulation (all 11 ACh levels)

The key experiment: does Model B transfer to real Allen Neuropixels
recordings better than Model A?
"""