Deriving a reporting cadence
From the original End
18 November 2026. The programme-decision foray asks for a small tool that identifies realistic information-seeking and a cadence matching project change. Earlier essais supplied delayed-observation mechanisms and policy heuristics. This one makes the timing of a paid observation a control decision and solves a finite benchmark exactly.
Powell’s Sequential Decision Analytics and Modeling supplies the state/decision/information/transition/objective discipline. His On State Variables, Bandit Problems and POMDPs clarifies the role of beliefs and learning decisions. The numerical model and recursion below are ours. The original climate default remains in its earlier experiments; the maintained festival-shuttle scenario is used here. No neural timescale is converted into a management cadence.
State and timing
The decision state is (t,p), with t=0…7 and p the conditional probability of present disruption using only received reports and the known transition law. The latent binary regime flips with probability h between consecutive windows. Predictive belief is T(p)=h+(1−2h)p. At each window: choose whether to pay for a perfect immediate observation; choose continue or protect using information now available; incur the window’s cost; transition to the next regime.
Continuing costs L[t] if disruption is present and zero otherwise. Protection costs C=3 in either regime and prevents that window’s disruption cost. It does not change the regime. The query fee is k. Unobserved incurred costs become known only after the horizon, so they do not provide a free sensor. There are no irreversible commitments, switching costs, capacity constraints or report delays in this benchmark. The latent process and costs are authored, not estimated.
Exact recursion and comparator
g(t,p) = min(p L[t], C)
V(8,p) = 0
skip = g(t,p) + V(t+1,T(p))
check = k + (1−p)[g(t,0)+V(t+1,T(0))]
+ p[g(t,1)+V(t+1,T(1))]
V(t,p) = min(skip,check)A tie skips the check. The operating action continues on a tie. Since the horizon and observation alphabet are finite, recursion over reachable beliefs is exact up to floating-point arithmetic. The solver never receives the realised hidden regime except through a paid report. No fitted sample, hindsight policy or privileged forecast is used.
For a fixed calendar, replace the final minimum by its preselected check/skip decision at each time. Enumerating all 2^8=256 calendars gives the best fixed comparator, including never and always. Operating choices still react to every receipt in every comparator. The adaptive policy may also change when it asks next. Costs include every chosen report once.
Results
Opening case: h=.1, k=.75, prior .2 and L=[2,2,2,12,12,2,2,12]. Optimal adaptive expected cost is 8.75686544; best fixed calendar 8.85608192 (windows 4 and 8); never check 12.18210560; check every window 12.66392192. The optimal policy always asks at window 4; later checking depends on the received state. This is a small adaptivity gain but a distinct solved question, and much of the gain comes from asking only where action can matter.
With every loss equal to 2, continuing is optimal in both regimes because protection costs 3. Paid information cannot improve any action; the optimal policy never pays for it, even if regimes switch quickly. Free information may tie and is omitted by the tie rule. This supplies a counterexample to cadence determined by rate of change alone.
Verification and limits
An independent oracle represents all 256 hidden trajectories and directly optimises query/skip branches using their unnormalised masses, rather than the model’s belief recursion. A parameter sweep compares exact total costs, including degenerate hazards, priors, free checks and prohibitively costly checks. Every fixed schedule is evaluated on the same process. This is numerical verification of the declared model, not practical project validation.
The UI’s reachable-state table and observation probabilities are ephemeral computations. Inputs, outputs and settings are not stored; reload resets them. There is no save or file workflow. Delayed/noisy sensing, uncertain transition parameters and irreversible commitments are the substantive next extensions: they can change both the information state and the optimal cadence, rather than merely restyle this example.