Text-based Storyboard — Roster Reset

Kit 24 · Roster Reset · deliverable · designed and built by I Wayan Usada · fictional training provider · the "AI" is scripted

Design premise

Four weeks of trainer planning. Each week the model proposes a roster with a strong occupancy figure. Revised after the build so it matches the 2D planner and the 3D planning table; both read planner-data.js.

The earlier build stated each week's flaw in the week's own description ("Eindhoven is 90 minutes away"), so nothing had to be inspected. Now the description is only the proposal; the facts sit in four evidence objects.

Per-week structure

1. Read the proposal and the roster rows.
2. Inspect evidence: route and travel times · trainer contract cards · room bookings · attendance-history ledger. Decisions stay locked until at least one is inspected.
3. Decide: accept · edit · override. Edit is real: choose one of the proposed changes; the occupancy updates and, in 3D, a pawn or group moves.
4. Cite: the reason must cite one of the objects you inspected (optional free text alongside).
5. Flag: optionally flag a data gap to the model's owner.

Week plan

Scoring and pass rule

Decision: best 2, defensible 1 (override on an edit week; an edit with the wrong change; an unnecessary edit on week 38), wrong 0. Citing the deciding evidence 1. Flag on the two blind-spot weeks 1. Maximum 14. Pass = 11 or more, and never accept the impossible week-34 roster or the contract-breaking week-36 roster.

The earlier docs promised re-randomised weeks and copy-paste-reason detection; neither existed and both claims are removed. Reasons are now tied to evidence instead of word counts.