Measure fact drift across repeated compaction

The experiment has three moving parts: a transcript with a known answer key, a compaction loop, and a scorer that asks the same 20 questions after every round. Every model call goes through one module that runs claude -p on your subscription, with the long context on stdin and no tools. The decision that matters most is the scorer's four outcomes. "Wrong" and "missing" are both losses, but a superseded answer (the old ground rent, the old completion date) is the one an agent states confidently, so it gets its own column.