Put one durable constraint in the system prompt, bury it under unrelated turns, then ask the model to do a task that the constraint governs. When it breaks the rule, fork the same conversation and ask it to list its active instructions instead. It recites the rule it just broke. Across seven pinned models and 875 scored outputs, 70 of 79 failures were that shape — 88.6% — which means the problem is not memory. Restating the rule as checkable predicates at every turn and verifying the answer against them took the weakest model from 54.0% to 78.4%.