Programme decision sequences · Essai 13 · 18 November 2026
When is another report worth asking for?
Choose the reporting cadence as part of the decision policy. A rapidly changing project need not justify a report when nothing valuable can be done with it.
Eight preparation decisions
The mountain-festival shuttle has eight preparation windows. An unobserved disruption switches between absent and present. At each window the organiser can continue normally, or pay three value points for protection that prevents that window’s disruption loss. Protection does not change future disruption. A perfect current-status check has a price, and arrives before that window’s operating choice.
Losses are scored after the eight windows: the organiser does not learn the regime from immediate losses. Between windows, either regime flips with the probability below. The transition law is known. These are deliberate finite-model assumptions, not calibrated transport forecasts.
Change the decision problem
Changes recalculate immediately; nothing is saved or transmitted. Reload restores the opening case.
| Observation rule | Expected cost |
|---|
Where the optimal policy asks
Bars are the probability of buying a check at each window, integrated over every possible report history. They are not eight independently chosen frequencies.
Inspect every reachable information state
Rows with equal time and belief are combined. Current belief uses received checks and declared transitions only. The two continuation costs already include future optimal decisions.
| Window | Chance reached | Belief of disruption | Cost if check | Cost if skip | Choice |
|---|
The gain, and its limit
The preceding essais showed delayed reporting and compared supplied review heuristics. This one solves the information-acquisition decision jointly with the operating decision. It evaluates every fixed calendar (256 schedules), then compares the best with a policy allowed to change its next query after a receipt. The same uncertainty law and cost account govern all policies.
Set every window to low stakes: because protection costs more than the worst loss, the operating choice never changes. Even a rapidly switching state then has no decision value worth paying for. Restore the uneven stakes: useful checks concentrate around actionable consequences, with future use of information included. The clock alone is insufficient.
This is an exact small benchmark under perfect, immediate checks. Delayed and noisy evidence requires a richer information state; the older temporal experiments retain that enquiry. It does not turn eight fictional windows into a recommended meeting schedule.
Derivation, source boundaries and independent verification · Model source