This probably is overthinking a bit but - is there a way to merge multiple color correction models?

My thought is create a separate tool (or integrate into this) that scans a video file for 4 frames - the frame with the most BLUE, the frame with the most RED, the frame with the most GREEN, and the frame where each is represented most equally. Output those untouched frames as png files.

Find the corresponding frames in what you're looking to match.

Create model for BLUE pair, GREEN pair, RED pair, and EQUAL pair.

Then create R, G, and B Predictions for Equal frame. Generate color models for untouched Equal to R-Equal Prediction, (and for BLUE and GREEN).

Then take the Predictions from those three, and average them with the original Equal to Equal Model.

While not as accurate as individual shot correction, I'd imagine the average of those 4 prediction models would give you the most accurate blanket color correction to apply to an entire film.