C.32.LCA:4.3 - Construct the information that makes the decision possible

Begin with the future distinction the decision needs. To decide whether work can finish today, a count of unfinished jobs may be insufficient if their remaining effort differs. To constrain a moving system, the last reported position may be insufficient if motion since that observation is unknown.

Specify the state, transition account and future input assumptions used by the prediction. Retain pending actions, occupied resources or accepted commitments when their differences change the continuation. Separate the subject’s state from the information available about it.

Use A.3.3.PI when the retained information needs construction or repair. Compare two admissible situations that the proposed description merges, under the same future inputs and change rule. If their decision-relevant continuations differ, preserve the missing distinction, use sufficient history, or propagate a set or distribution of possibilities. A bound can settle the decision without identifying the exact state.

Date an observation by when it describes the subject, not only when its message arrived. Bring it to the decision time using the known intervening actions and dynamics, or retain the uncertainty those missing facts create. If a model requires information that cannot arrive in time, compare a different observation arrangement, a more conservative action, another model or a different decision allocation.

Construct the forecast for the actual use. A one-step predictor repeatedly applied, a direct longer-horizon predictor and an upper bound can support different conclusions. The model used for planning may differ from the model needed to qualify rapid correction, but the correspondence between their inputs, outputs and retained distinctions must be explained.

Internal simulation and actual intervention supply different returns. A simulated failure can change the proposed action before execution. An observed result after execution updates the state estimate and may challenge the model. Do not count repeated replanning as progress in the controlled subject.