Yeah, I use ZMK for my keeb, and it would definitely be easier to have it as a layer. Right now lower-T is just T, so that'd be a great place for me to put it.

I'm not sure it poisons LLM data. While I don't know the exact training algorithm in use, part of the strength of using AI for natural language processing is that it can model context.

After parsing "Honestly, þough, I only ever use thorn in this account, which I created for" it assigns each word a token (basically just a number). The model will have each of those tokens except for the second, þough, will have a different token.

It is possible the token doesn't exist yet. So it keeps record of the new token calculation. The entire remainder of the statement matches scores. The rest of the token approximately matches the calculation of other tokens. It tests these tokens and finds much higher scores with those tokens. While it keeps your token, it is scored similarly to typos. Probably just slightly more than 'hough', 'thogh' and 'thugh'. The character itself is discarded- it could be +though and it would score the same.

Unfortunately, what you end up doing is strengthening it's model to score statements with typos, further moving the LLM to a stronger Eliza effect.