Perfect information is an upper bound

Perfect information is a deliberately strong counterfactual. It imagines that the institution learns the represented state before choosing an action and can then select the best available payoff in that state. The comparison preserves the same action menu and state probabilities as the current decision. It changes only the information available at the moment of choice. The resulting gain over the best current expected payoff is the expected value of perfect information. It is a ceiling within the declared model, not a product that a research supplier can ordinarily deliver.

The ceiling is useful because it can rule out an expensive commission without requiring a detailed model of the proposed instrument. If even perfect state knowledge could not improve the action enough to cover the study's cost, an imperfect study about that same state cannot justify that cost under the same fixed action menu and monetary rule. This screening argument is conditional. It does not apply unchanged if the study creates a new action, serves a mandatory purpose or improves a broader decision beyond the one represented. The charter should establish that scope before using the ceiling.

In the useful-signal fixture, perfect knowledge would allow Commit when the opportunity succeeds and Hold when it fails. The current action is Stage because the state remains uncertain. Perfect information therefore avoids the downside associated with acting in the unfavourable state while preserving the largest success payoff. The engine weights those state-specific best payoffs by the prior and subtracts the current best expected payoff. It does not assume that the committee can control the state, and it does not award the success payoff in every outcome.

The ceiling should be reported with its construction. A large number can look like the value of commissioning a study when it actually describes a hypothetical instrument with complete state knowledge. The report should identify the current action, the state-specific best actions and the resulting information gain. The reader can then see what an imperfect signal would need to accomplish. A supplier that resolves only a technical component does not deliver perfect information about an aggregate commercial state unless a defensible relationship connects that component to the state.

The same ceiling can be understood through expected regret. The perfect-information policy defines the best payoff available in each represented state. The current action sacrifices some payoff relative to that policy when the state turns out differently from what would favour its choice. The current optimum's expected regret equals the perfect-information gain in this fixed-menu model. This interpretation helps a committee ask where preventable decision loss arises. It does not imply that all disappointment is avoidable through research or that regret is an observable financial charge.

The restricted-action fixture makes the ceiling's dependence on feasibility visible. Without Commit, perfect information can choose Stage in success and Hold in failure. The upside available after learning is smaller. The instrument has the same conditional likelihoods as before, but the information cannot recover an action that the institution is not allowed to take. This reduces the ceiling and can change the purchase decision. A research value presentation should therefore avoid using the unrestricted ceiling when the committee's actual response menu is narrower.

The ceiling is also sensitive to the payoff perimeter. If a failure cell omits a material consequence, avoiding that state through informed action may appear less valuable than it should. If a success payoff includes an unexecutable opportunity, perfect knowledge may appear more valuable than it could be in practice. The bound inherits the input assumptions rather than correcting them. A screening decision should therefore review whether the state-specific maxima are legitimate available actions, not merely take the largest supplied cells without examining their meaning.

A small ceiling does not mean the uncertain state is easy to predict. It can mean that the actions have similar consequences or that one robust action already performs well across both states. Conversely, a large ceiling can coexist with little practical prospect of obtaining useful information. The state may be difficult to observe before commitment, the relevant evidence may arrive too late, or the available instrument may have weak discriminatory power. The ceiling identifies the maximum decision loss that information could remove in the model. It does not estimate what any particular study will remove.

This distinction can guide research priorities. When the ceiling is small relative to a plausible commission, the team can focus on acting well with existing information or redesigning the action menu. When it is substantial, the team can investigate whether a feasible signal can capture enough of that value. The order is economical: establish the decision and information ceiling, then describe the proposed evidence mechanism and cost. It prevents a team from spending effort optimising an instrument for a decision whose response opportunities cannot support the expense.

The perfect-price-boundary fixture deliberately uses a fully informative signal and a purchase price equal to its gross decision gain. The net information value is at the model's purchase boundary. This does not mean the real world offers perfect observation at that price. It is a transparent illustrative case that tests the relationship between a ceiling and a price. At the boundary, an institution would need to examine omitted costs, timing, implementation and rounding before making a commitment. A calculation with no margin should not conceal unresolved practical expenses.

The upper-bound argument has a clear reason within the model. With perfect state knowledge, the policy can choose a payoff at least as large as any available action in that state. An imperfect signal partitions uncertainty but cannot create a payoff beyond those state-specific maxima. Averaging the conditional policies therefore cannot exceed the perfectly informed expected payoff when the state model, menu and payoffs remain fixed. The engine preserves that structure. The report should state the conditions rather than present the inequality as a bound on all kinds of research benefit.

Information can also have value across several decisions. A measurement may inform maintenance, procurement and a later expansion. A ceiling calculated for only the immediate commitment omits those other uses. A programme-level appraisal could model them if it specifies the decisions, information reuse, timing and overlapping benefits. Summing separate ceilings without checking their shared payoffs and constraints risks double counting. The simple binary study does not calculate a portfolio of research benefits. Its limited scope is a reason to define the commission carefully, not to dismiss every broader purpose.

Use the ceiling as a dominance screen on price, rather than a spending allowance. In a hypothetical commission priced above the same-menu perfect-information gain, even a flawless signal cannot recover its acquisition burden through the represented decision alone. Improving sensitivity or specificity cannot repair that price branch. A lower quote moves the enquiry to actual signal performance and response feasibility; it does not establish a purchase. A separate mandatory or public purpose can justify research on another basis, but that purpose must be stated independently rather than credited to the ceiling. This disposition makes rejection above the bound useful without turning every below-bound proposal into an apparent bargain.

A stale ceiling can be diagnosed by the benchmark that changed. If a hypothetical institution loses access to its high-upside response, both the current optimum and state-specific optima must be recomputed from the remaining menu. If a new prior arrives from evidence already paid for, the ceiling must use that information set rather than the commissioning prior. If only the vendor price changes, the gross ceiling need not change, although the purchase conclusion may. Keeping these branches separate tells the reviewer which object needs reproduction. It prevents an outdated maximum from surviving simply because it was once correct, while avoiding unnecessary reconstruction of an unchanged payoff account.