Appendix: probability and value dictionary

The reproducibility dictionary connects a decision term to a specific computational field and meaning. Its purpose is to prevent a reader from treating similar labels as interchangeable. Prior probability, instrument sensitivity, specificity, branch mass and posterior success probability are all probabilistic quantities, but they condition on different information. Expected payoff, weighted branch contribution, gross information value and net information value are monetary quantities with different perimeters. A complete reproduction object preserves these distinctions rather than relying on a summary metric alone.

The input fixture is the supplied decision problem. It contains a prior, conditional signal likelihoods, an aggregate study cost and the action payoff cells. The fixtures here are explicitly illustrative. None is a fitted empirical object. The engine validates the input structure, applies its declared binary risk-neutral method and returns a structured output. The claim ledger then identifies each scalar selected for publication. A reader should interpret a displayed number through that chain, not as a free-standing fact whose meaning survives every change of context.

The prior field is displayed in percentage units on input, while branch mass and posterior values are stored as fractions on output. The binding format can display those fractions as percentages. This is a presentation transformation, not a revised belief. The full JSON remains the authoritative computational object. A reader reproducing a figure should use its selected raw output pointers and the stated plot scale. Multiplying a percentage input again as though it were a fraction would change the model, so units belong in both the dictionary and the exhibit.

The action cells are aggregate monetary payoffs in USDk. The app does not infer a duration, exchange rate, discount schedule or cost incidence from that label. Those meanings must be supplied through the payoff perimeter. The same applies to testCost: the engine subtracts the aggregate amount once from gross information value but does not assemble its vendor, internal, waiting or operating components. An actual dossier should retain the separate evidence behind that aggregate. The fixture's transparent simplicity is not an empirical certification of every input component.

The actions output is ranked by current expected payoff, not preserved in the original input order. The regret output preserves a separate action account. A scalar pointer therefore needs to identify the actual structured location and its meaning; an author should not assume that a row index always names the same action in every output. The paper's binding files select values from the named fixture and the tables show action readings alongside them. If an input changes the ranking, the output and narrative must be checked again.

The branches output separates the signal name, probability mass, posterior, chosen action, conditional expected payoff and weighted contribution. The conditional payoff is evaluated after the signal. Its weighted contribution enters the ex ante expected payoff with information. These are not additive without the branch weights. A rare branch can have an attractive conditional value and a smaller contribution to the full commission. The dictionary makes the weighting step visible so that a favourable branch is not presented as though it were the certain payoff of buying a study.

The perfect-information account uses the state-specific best payoffs from the same menu and the same prior. It is not a stronger estimate of the actual study's performance. The resulting ceiling can screen a proposed purchase under the declared scope, while the signal-specific gross gain requires the branch policy. Both subtract the same current best expected payoff. A reader should inspect that benchmark when comparing the figures. If different values or action menus have been used, the apparent relationship between gain and ceiling may no longer describe one coherent decision.

Net information value subtracts the supplied study cost from gross signal value. A positive result is a conditional purchase finding under the stated risk-neutral objective, not an authorisation of the eventual operating action. A negative result can coexist with a useful instrument or a legitimate broader research purpose. The dictionary should retain those distinctions in the field descriptions. The summary metric becomes more useful when the reader knows exactly which part of the institution's commissioning question it answers.

The regret fields compare each action with the state-specific best available payoff. Expected regret uses the same prior weighting. The app does not convert regret into a psychological scale, determine a maximum-regret policy or elicit utility preferences. Publishing those labels without their scope could suggest features that the engine lacks. The table here identifies them as outputs of the fixed monetary matrix, and the conceptual preference chapter describes the additional methods that might be needed for a real institutional rule.

The priorSweep output is a sampled current-policy account. It varies the prior through the engine's predefined points and reports the best action and value at each point. It does not vary signal performance, calculate EVSI at every point or assign probabilities to alternative priors. The thresholds output contains pairwise equality points plus the complete-menu best action at each point. These are candidate comparisons under linear state payoffs. The dictionary prevents the plot from being mislabelled as a continuous signal-value curve or every pairwise point from being called a policy transition.

The complete output contains display-oriented metrics, tables, warnings and method descriptions alongside decision-relevant values and the exact rational evaluation. These fields support the public app interface. The paper's quantitative claims bind to the named fixture and corresponding domain values. Duplicate representations of a number in a display table are not independent evidence. If a narrative relies on a qualitative warning, check whether its stated conditions hold in that fixture. In the no-policy-change case, exact EVSI is zero because the implemented policy remains unchanged. The stored exact fractions establish that arithmetic identity; converted numeric displays support reading without creating a different computational result.

The bindings file supplies the bridge from a published claim token to its input or output pointer. It retains the basis, model owner and associated ledger entry. A displayed claim should have reciprocal references to the block in which it is used. A selected plot coordinate needs the same link even when its value is not printed in the caption. This makes the figure reproducible from the declared object. It also prevents a chart from silently using a calculation different from the table beside it.

The dictionary's final category is unimplemented method. Causal estimation, utility elicitation, joint evidence, dated delay, adaptive stopping, implementation performance and research programme valuation are discussed where relevant. They are not domain outputs of the binary app. A real dossier needing one of them should specify its method and evidence before making a numerical claim. The presence of a chapter or worksheet does not substitute for a computation. This boundary is part of the reproduction object because it states what the object cannot honestly demonstrate.

Use the dictionary to trace a suspected interpretation error to its conditioning event. A hypothetical reviewer comparing the positive branch posterior with the unconditional prior is comparing different quantities; one is conditioned on the signal and the other is not. Comparing branch mass with sensitivity also mixes a marginal signal probability with a state-conditional likelihood. Comparing weighted branch contribution with conditional expected payoff omits the branch's occurrence mass. These errors can survive immaculate units and accurate copying. The field map should identify the conditioning statement before a value is transferred to prose, a chart or a response rule, then retain the complete input and output identity for the comparison.