When there are a lot of MC options, forcing min of 1% makes things error

[CodexVeritas]

For example here. Keeping a min of 1% makes sense for binaries, and most MC, but we should probably scale the min based on the number of options (or remove the min for MC). The LLM will give a prediction of 0% from time to time.

Example problem question here

[CodexVeritas]

Another example question and workflow