# Constraint-planning record

This hypothetical teaching record describes a planning decision. It grants no permission to assign work. The rates are authored USD accounting inputs, not measured expenditure or savings.

## Completed example

- Work: one metric report requiring analytics and one access check requiring support.
- Duration: each task takes exactly one hour. Both must start at 09:00.
- Candidate slots: 09:00 and 10:00 on one hypothetical day. Task constraints still permit only 09:00.
- Capacity: one task per worker in a slot; each task assigned exactly once.
- Avery: analytics/support; available both slots; $80 per hour.
- Blair: analytics; available 09:00; $60 per hour.
- Casey: support; available 09:00; $70 per hour.
- Objective: minimize total declared assignment cost; operational owner must approve its suitability in a real workflow.
- Candidate space: (3 workers × 2 slots)² = 36 assignments.
- Feasible schedules: report Avery/access Casey $150; report Blair/access Avery $140; report Blair/access Casey $130. All start 09:00.
- Minimum: $130 in this declared space. No actual cash saving is established.
- Invalid proposal: Avery on both tasks at 09:00. Reject as double-booked even though both skills match.
- Missing-skill variant: no feasible plan; best cost null. Escalate to the owner rather than silently changing the skill or deadline.
- Tie variant: equal $70 rates produce three best plans at $140. Preserve the tie; no fairness claim follows from cost equality.
- Runtime boundary: offline enumeration, at most 50,000 candidates. No model call, dispatch, production permission check, live input refresh, or scale claim.
- Local checks: assignment completeness/uniqueness, worker identity, skills, allowed slots, availability, capacity, objective minimum, ties, no-solution behavior, and rejected proposal rule overrides.
- Target evidence still missing: observed workload, owned rules, duration variability, calendars/timezones, current data, permission and approval checks, solver budget, concurrency, source-of-truth readback, and operator acceptance.

## Your working record

1. What decision will this plan support, and who owns it?
2. What individual work items must be assigned, and how long does each take?
3. Which source and revision establish skills, availability, deadlines, and capacity?
4. Which constraints are mandatory? Which are preferences? Who may change each?
5. What objective should be minimized or maximized, in which units? What gets excluded?
6. Can a rule or direct assignment meet the need? Why compare a solver?
7. Which known bad plans must fail? What proves the validator catches them?
8. What does a tie mean? What happens if no feasible plan exists?
9. What input changes require revalidation before committing the plan?
10. Who may approve and execute it? Which trusted boundary rechecks authority and verifies the result?
11. What solve-time/resource budget, fallback, operating owner, and retirement rule apply?
12. What evidence supports selecting a mechanism? What remains unknown?

Decision: investigate / test a bounded candidate / keep the current method / defer.

Owner and review date: