5) a computational model beautifully done by Guanchun Li that implemented BTSP learning rules and a synaptic weight consolidation mechanism was able to generalize and made some mechanistic predictions

check it out!

(continued) task-related and context-related information were encoded in separate subspace (see Wenbo Tang's preprint). Thus, it is active generalization, not passive reward-distance coding transfer. Moreover, this generalization process allows faster adaptation to reward switching in the future.

3) weak residual activity, influenced by past learning ("memory traces") and presumably driven by CA3, biased the BTSP events to guide new learning for generalization.
4) place cells changed reference frames from spatial cues to reward goal when switching contexts from familiar to novel belts.

Yes super helpful. I stick to the context-content-conclusion rule for every main figure paragraph.

You can use colorbrewer to start with. It sounds suspicious to input your figure and output a color-blind-friendly version without distorting anything.