Spent my evening after work playing with Coldstore's Chaos language. While it already works, it's all direct interpretation, and only minimal error checking for stack depth requirements. So I spent the evening extending the primatives list to include stack information.
Next will be some decisions on how to do bytecode representation, and a seperate version of the lexer that emits the bytecode, rather then direct interpretation.
Looked at a RPN based functional language called Joy -- there's definitely some interesting concepts that could be borrowed from it, and it's one of the reasons for adding the stack effects notation to the current set of primative operations. I think in a later iteration, I'll try to use that to guess 'correctness' of procedures and stacks.