How do you handle coverage guarantees from split conformal prediction when the data shifts? [D]

Split conformal prediction gives you a clean statement: with a calibration set exchangeable with future data, your prediction sets cover the true label at the rate you chose. The trouble is that production data rarely stays exchangeable. Categories get added, ticket wording changes, and a calibration set from three months ago quietly stops being representative.

I’m interested in how people deal with this in practice. Do you recalibrate on a schedule, monitor empirical coverage and alert when it drops, or use an adaptive variant? And how do you decide what to do with the cases where the set contains several labels, such as route them to a human, an LLM, or a fallback rule?