Scales encode value judgments

A common scale makes attributes arithmetically comparable by choosing what differences mean. It does not discover that meaning in the data. Capital expenditure, a strategic-impact judgement and a readiness judgement arrive in different units and with different interpretations. Turning them into a common range allows addition, but the transformation determines how much an observed difference contributes to the total. The scale choice therefore belongs in the preference architecture. It should be explained before the committee interprets a score gap as evidence of a better commitment.

The canonical app uses an observed-range transformation. For an attribute to be maximised, the least favourable supplied value receives the low end of the common scale and the most favourable supplied value receives the high end. Intermediate values are placed linearly between those extremes. Cost and execution risk reverse the direction because lower supplied values are preferred. When every candidate has the same value on an attribute, the app assigns a common midpoint. This is a specified calculation convention. It makes no claim that the candidate extremes represent an absolute failure or ideal outcome.

The convention has a useful property for exploratory comparison. It preserves the direction of raw preference and displays how each alternative sits relative to the available set. It also prevents currency amounts from overwhelming a dimension simply because their units are numerically larger than an index. Those benefits are conditional on the raw values being comparable and the linear transformation being an acceptable representation of value. A transformation cannot repair inconsistent cost scopes or establish that equal changes on an ordinal judgement carry equal preference value.

Linearity is a substantive assumption. The organisation might value the first improvement in reliability more than a later improvement after service is already acceptable. It might place little value on readiness below an executable threshold and substantial value on reaching that threshold. An impact rubric may describe qualitative categories whose numerical spacing is only a convenience. The app does not estimate a nonlinear value function or infer these thresholds. A user who adopts its linear transformation should understand what the approximation permits and where it may distort the intended tradeoff.

Observed ranges also give the candidate universe control over the scale. If the best and worst supplied values move, the same raw difference can represent a different portion of the normalised range. This is not an implementation error. It is how the chosen transformation works. It becomes a decision error when the organisation compares scores from different sets as if their scale had remained fixed, or retains weights elicited for a different swing in performance. The interpretation of the weight and the interpretation of the scale must remain connected.

The set-extension experiment is designed to isolate that dependence. Its starting point is the balanced illustrative mandate. The extension retains every existing alternative, raw criterion value and declared weight, then adds an expressly inferior reference proposal. The added proposal has weaker impact and readiness, greater cost and worse execution risk than the existing candidates. The experiment is not an argument for admitting such a proposal. It asks what the canonical observed-range engine does if the universe contains it, so that the governance consequence becomes visible.

The important comparison is between the unchanged alternatives, rather than between the new proposal and a winner. If their raw performance and weights stay fixed while their preference scores change, the change must be interpreted through the new scale. The experiment therefore retains both the input record and the computed output. It does not describe a higher score as improved operating performance. It does not describe the added proposal as a newly available market option. Every supplied quantity remains explicitly illustrative and belongs only to the frozen study.

The balanced illustrative set gives Reliability architecture a score of 72.13 preference points. With the dominated proposal included, the same alternative receives 89.98 points and remains the displayed leader. The added proposal receives 0.00 points. This experiment demonstrates changed score levels while explicitly disclosing the absence of a leader reversal in these supplied cases. It cannot establish how frequently set changes would reverse a recommendation in any population of investment decisions.

The result changes the interpretation of a presentation. A sponsor who reports only the extended-set score might make an unchanged proposal appear more compelling than it looked under the original set. The higher number would be correct for the new calculation, yet misleading as evidence of improvement. The repair is to show raw values, name the universe and explain the transformation. A committee should be able to see whether it is discussing a better intervention or a different yardstick. The score cannot supply that distinction without the input record.

Fixed external anchors are a conceptual alternative. The organisation could define a relevant worst and best performance level independently of the candidates, then place each candidate against those anchors. That approach can improve continuity across candidate sets, but it creates other responsibilities. The anchors need a legitimate objective basis, an evidence perimeter and a review rule when the operating context changes. Anchors chosen merely to favour a proposal are no more defensible than a manipulated observed range. This paper describes the alternative architecture; the current app does not implement it.

A value function adds another distinction. It represents the committee's preference over performance levels, rather than merely rescaling an observed range. A function can express saturation, a threshold or an asymmetric concern if those features belong to the authorised preferences. It should not be fitted casually to make an inconvenient candidate lose. Eliciting and validating such a function requires a separate method and record. The canonical linear app is useful partly because its simplicity exposes where a richer preference model would be needed. It should not be presented as if it had already estimated that richer model.

A scaling review can therefore be a small set of decision questions. What raw difference does the full normalised swing represent? Is an equal increment in the raw value an acceptable equal increment in preference? Would adding or removing an irrelevant candidate materially change the interpretation? Are the supplied weights intended for this specific swing? Does a required threshold belong outside the compensatory score? The answers determine whether the observed-range convention is appropriate for the current purpose, rather than whether the mathematics can execute it.

The organisation should retain a scale version with the decision record. If the candidate universe changes, the version should change even when the raw source files for the existing alternatives do not. If a criterion definition changes, the team should revisit the raw values as well as the transformation. A later reviewer can then distinguish an updated operating assumption from a changed value convention. This discipline supports learning without treating every historical score as though it belonged to a single permanent scale.