Hallucination as output-boundary misclassification: a composite abstention architecture for language models

Build a pluggable 'abstention gate' library that allows any black-box LLM to selectively refuse to answer based on a composite support deficit score. This layer sits between the model and the final output to reduce hallucination.

Suggested repo: GateKeeper-LLM

"Force your LLM to say 'I don't know' when it's just guessing."

Estimated effort: 35h