Goal:
Repair TT639B's new simple-QA/assistant/explanation slips while preserving:
TT638D dense + dyadic proof lineage
TT639B regression repair
TT639B rule/evidence behavior
This is still a tiny exact assistant ladder, not a broad chatbot/general-knowledge claim.
Evidence:
TT639B passed regression + rule, but failed seen/heldout in simple groups such as assistant_role, tests_matter, yes/no reasoning, return_statement, and capital_france.… See the full description on the dataset page:
https://huggingface.co/datasets/CircularBalls/tt639c-rebalanced-assistant-ladder-v1.
