Brian, that retrieval-drift line is the whole game. Demos run on frozen context; production drifts weekly, and the failure shows up three hops downstream where nobody is looking.
That is the bet behind receipts: when every tool call is logged with its inputs and an output hash, the eval loop stops being a separate harness you run on Fridays and starts being a byproduct of normal operation. You don't discover the drift in a retro. You see exactly which calls started failing, and when.
Curious what bit you earliest: retrieval quality, or something dumber, like a tool silently returning stale results?