The current mapping was derived as follows:

1. Download en-US and en-GB wordlists from http://wordlist.aspell.net/
2. Use a script to compare these lists, and build us_only.txt and gb_only.txt.
3. Heuristically match words from the two lists with similar spellings
4. Using this match list as a shortlist of likely candidates, manually create a one-way substitution mapping (en-US->en-GB)

The heuristic matching was done by trying the following spelling substitutions on US-only words and seeing if they match any GB-only word:

re.sub(r'ol', 'oul', us_word)
re.sub(r'or', 'our', us_word)
re.sub(r'er', 're', us_word)
re.sub(r'([lprst])(ing|ed|ers?)$', r'\1\1\2', us_word)
re.sub(r'se', 'ce', us_word)
re.sub(r'yz', 'ys', us_word)
re.sub(r'sk', 'sc', us_word)
re.sub(r'e(?=[acdgmos])', 'ae', us_word)
re.sub(r'e(?=[abs])', 'oe', us_word)
re.sub(r'll', 'l', us_word)
re.sub(r'og', 'ogue', us_word)

We might benefit from an initial small set of words that we can take to production which won't be too controversial (or difficult to review) while using the full list for feedback.

I have been considering whether it's worth sorting our proposed list by frequency to identify a smaller "starter" list, but it's unlikely to be worth the effort. I had a quick look around and found that the dataset on this github repo, which is licensed CC-by-sa-4.0. I believe this would be a licence-compatible dataset to use. However, we don't have a mechanism to attribute in the configuration files and I'm not sure if attributing in the commit would be sufficient.

Arguably a quick once-over for words I recognise, or a quick sort by length, might be a more efficient use of time. I'll collect feedback.

For future consideration, Wikidata has lexemes (e.g. color/colour) that note varieties of English. It might be possible (but perhaps difficult) to generate a list of conversions based on those.