A new paper from the School of Emergent Priorities reports fourteen months of an agentic intake tool that forecasts reclaim odds and schedules disposal dates for the University's lost-property store, and finds its human-review checkpoint barely alters the tool's own decisions.
A fourteen-month deployment across six collection points lifts reclaim rates, while a review checkpoint staff open often changes almost nothing
An item used to wait for someone with a free afternoon. Now it waits for a date the system already set for it at the door.
— Dr Anouk Mensah, Postdoctoral Fellow in the School of Emergent Priorities
That system, the Lost Property Ledger, has run at six of the University’s staffed collection points for fourteen months, replacing the paper logbook that used to sit behind the library desk and its counterparts across campus. A new paper from Dr Mensah, Associate Professor Casimir Beng, and Dr Mirela Hanke reports what changed once an item’s fate stopped waiting on a reviewer’s spare hour: a photograph and a short description at intake now produce a reclaim-probability forecast, a retention tier, and a disposal date, without anyone drafting a decision.
Reclaim rates rose at five of the six collection points, and a backlog that had nearly quadrupled in the fourteen months before rollout levelled off rather than continuing to climb. A checkpoint letting staff reopen and amend any forecast before its scheduled date saw heavy use — opened on a third of eligible items — but changed the tier in barely one case in forty of those openings, a gap the paper’s authors read as worth a closer look in its own right rather than a flaw to correct. The School treats a finding this modest in scale, and this unglamorous in subject, as no smaller a return on the study than a dramatic one would have been.
“We built the checkpoint assuming it would be exercised,” said Associate Professor Beng, Lead of the Adaptive Metrics Lab. “What we got instead was a number worth reporting on: how often a step gets opened, against how often it does anything once it is.”
The full paper is available from the University’s research repository under an open licence, doi:10.5555/slop.7682cb.