Kit 24 · Roster Reset · deliverable · designed and built by I Wayan Usada · fictional case
In this fictional brief: planners using the AI rostering tool check proposals against the evidence the model cannot see, correct them in the tool, and report data gaps so proposals improve. Proposed measures are on the case page; nothing has been measured.
Accept high-occupancy proposals unread, or override everything by hand. Either way the model never learns.
| # | Behaviour | What it looks like | Where the build practises it |
|---|---|---|---|
| 1 | Inspect before deciding | route, contract, rooms or history opened first | decisions locked until an object is inspected |
| 2 | Choose the instrument | accept, edit or override, per week | decision scored 2 / 1 / 0 |
| 3 | Edit, not rebuild | the smallest change that fixes the week | real edit options with occupancy effect |
| 4 | Cite the evidence | the reason names what was inspected | cite step scored against the deciding evidence |
| 5 | Flag the data gap | travel time and history reported to the model owner | flag scored on weeks 34 and 40 |
It cannot add the missing data to the model. The case page recommends that to the tool owner.
Prefer the source? Download the raw markdown.