Programme Decision Sequences

New separate essay: 3 — When is evidence worth waiting for? · An exact decision tree for evidence cost, delay and commitment. The two original views remain below.

Purpose

Make the governance loop explicit

The manager does not choose an action after seeing the future shock. Information received since the previous choice is already represented in S_t. The current choice x_t is followed by new exogenous information W_t+1, which updates the next state.

This corrects the common slide-deck blur between “observe, decide, evolve” and Powell’s compact repeating sequence. The distinction matters when testing whether a policy is using only information available at decision time.

S_t · pre-decision stateProgramme position, last received signal, feasible actions and beliefs about drift and intervention effectiveness.

x_t · decisionChoose programme effort and information effort, respecting the quarterly ramp constraint.

W_t+1 · new informationClimate variation, measurement noise and execution error arrive after the choice.

S_t+1 · transitionUpdate temperature estimate, cost, schedule pressure and the belief covariance; repeat.

Model controls

Policy representative

Decision horizon · quarters

Temperature guardrail · illustrative units

Observation noise σ · per quarter

Reproducible path seed

All coefficients are deliberately stylised. Change them in the source before treating this as even a candidate programme model.

Final observed level—

Programme cost—

Model objective—

Inspect transition: quarter 0

Corrected taxonomy

Four policy classes, not four rival fields

These are small representatives of Powell’s classes, not claims of optimality. Reinforcement learning can supply methods inside or across these classes; it is not substituted for the fourth class.

PFA · Policy function approximation

A direct mapping from state to action. Here: a transparent set-point rule with an uncertainty-sensitive information choice.

CFA · Cost function approximation

Choose an action by optimising a deliberately modified immediate cost. Here: penalties add target margin and uncertainty reserve.

VFA · Value function approximation

Choose using immediate contribution plus an approximation of downstream value. Here: remaining effort, target gap and belief variance.

DLA · Direct lookahead approximation

Illustrative four-quarter rollouts—not MCTS or a proof of optimality. The tail policy uses sampled hidden parameters without learning inside the rollout, so this representative cannot establish the value of evidence. See the exact information-respecting essay.

P90 objective · lower is better

The objective is an explicit modelling choice: programme cost + schedule pressure + guardrail exceedance, with a terminal penalty. It is not a real climate-policy welfare function.

Class	Median final	P(above guardrail)	Median cost	P90 objective

Selected policy · path distribution

Scenario tree · not MCTS

The tree branches on possible information arrivals and displays the resulting policy decisions. It does not maintain visit counts, explore actions or back up values, so calling it MCTS would be incorrect.

Management framing tool

Match decision cadence to information cadence

Start with a minimum viable decision model. Use RL only when repeated interaction, a stable state/action representation and dependable feedback justify it.