A heat map encodes ordinal judgments

An ordinal heat map can communicate relative concern. It cannot acquire cardinal meaning merely because its categories are assigned numerals or multiplied. A severity category orders consequences under a chosen convention; it does not necessarily measure a ratio of losses. A likelihood category may order judgments without defining a probability interval. When those categories are combined, the resulting score reflects the classification rules. It does not automatically become annual expected loss, a comparable monetary burden or a sufficient basis for accepting an exposure.

The problem is not that qualitative judgment is illegitimate. Qualitative methods can be the honest choice when quantitative evidence is weak, consequences are diverse or the immediate purpose is to direct expert attention. A matrix becomes misleading when the institution forgets what its categories mean or uses the score for a decision the categories cannot support. If a report labels a risk low because its event is rare while its consequence could defeat a non-compensable objective, the reader needs the underlying consequence and acceptance condition. An attractive green cell cannot erase that information.

A useful triage score has a declared job. It might identify which risks require a specialist review, which actions need a higher authority level or which evidence gaps should be addressed first. Those jobs can be tested against cases. A classification rule that misses a severe consequence under the institution's own policy should be changed or supplemented. A rule that consistently directs attention to the right challenge may remain useful even though it is not a monetary measure. The institution should assess the score by the decisions it supports rather than by visual balance across a grid.

The paper's equal-expectation pair illustrates a related limit even for cardinal arithmetic. The dispatch source and rare-major case have matching expected-loss totals under their stipulated inputs and shared-control mechanism. Their event probabilities and event severities differ markedly. A decision based solely on annual expectation would treat their expected financial burdens alike. A decision concerned with the consequence of a single event could treat them differently. Neither disposition is inconsistent if the decision condition is explicit. The error would be to assume that equality of a mean proves equivalence of every consequence relevant to authority.

The rare-major fixture does not calculate a maximum lifetime loss or a tail percentile. It supplies a much larger loss per event with a smaller annual event probability. The canonical expected-loss mechanism preserves that distinction in its inputs and residual probability/severity outputs. A reader may examine those quantities without pretending a full loss distribution has been modelled. This is a useful stopping point: the simple model can identify that an expectation-only policy may be insufficient, while a specialist assessment determines whether the larger consequence is tolerable under the actual operating mandate.

The completed fictional disposition is to hold any claim that the matching expectations make the cases equally acceptable. The Model Risk Chair has no supplied consequence mandate for the rare-major case. The Consequence Reviewer must determine the relevant acceptance condition and identify whether the loss description captures the operating effects that matter. This is not an instruction to reject every rare large loss. It is a refusal to let the matching mean perform a decision that requires additional consequence and authority evidence.

That route demonstrates an advantage over both colour-led and expectation-led acceptance. The source mechanism may pass an optional annual net-value condition while the rare-major case remains held for a consequence review. The economic calculation is not discarded; its scope is retained. If the actual institution later supplies a permissible consequence condition and adequate evidence, the held branch may change. If it supplies a prohibited consequence, a positive expected economic result cannot override it. Both outcomes remain visible without adjusting the original model to make the committee's chosen policy appear inevitable.

Risk categories can still make that distinction visible if designed appropriately. A separate consequence constraint can route the rare-major case to a specialist authority regardless of its financial expectation. The score then performs triage rather than acceptance. Such a design should not hide its non-compensable rule inside an opaque weighted total. A very small likelihood contribution should not cancel a consequence that the institution has explicitly refused to accept. The condition should appear as its own decision branch with an identified owner and evidence requirement.

This is why a single composite risk number often asks too much of a reporting system. The institution may be trying to compare expected financial cost, continuity, safety, legal obligations, reputation, reversibility and evidence quality in the same scalar. These dimensions are not automatically commensurable. A weighted score can be useful when tradeoffs are consciously accepted, but it should disclose how one attribute compensates another. If compensation is prohibited, the rule should remain a constraint. A record that states the constraint is easier to challenge than a score that silently embeds it.

An ordinal ranking also depends on classification boundaries. Moving an input just across a category boundary can change a colour even when the underlying mechanism changes little. Conversely, a material change within the same category can leave the colour unchanged. Neither behaviour proves the matrix is wrong; it identifies its resolution. A responsible reader should know whether the classification is suitable for the decision's scale. A coarse category may be sufficient for an initial review but insufficient for choosing between close, expensive alternatives.

Evidence quality adds another dimension. A favourable numerical estimate based on a weak assumption should not necessarily outrank a more cautious but better supported estimate. The score and the confidence in its basis are different objects. An institution can report both without manufacturing a probability that the risk assessment is correct. The fiction uses statuses such as STIPULATED INPUT and VERIFIED ARITHMETIC to distinguish a supplied parameter from a checked calculation. It never treats a numerical output as empirical confidence in that parameter.

A matrix can make a useful conversation harder if its label becomes socially difficult to challenge. An Operating Owner may resist moving a risk to a more serious category because that would suggest poor management, even when the changed category simply reflects better information. A board may interpret a reduction in reported high risks as progress without asking whether controls were tested or classifications changed. The mechanism record counters that ambiguity by retaining what changed: the event definition, the evidence, the control effect, the cost, the reporting rule or the actual decision.

The same discipline applies to risk appetite. Appetite is not established by choosing a colour threshold and calling everything below it acceptable. The institution should name which consequences it is willing to bear, for which objectives, under what evidence and authority. An appetite statement may be qualitative and still operational if a material adverse scenario has a clear response. It may be quantitative and still vague if no one knows which loss, period or aggregation perimeter it concerns. Methodological precision should follow the commitment rather than the format of the statement.

The goal is to preserve useful compression without losing the reason for the next act. A board could receive a short table listing the decision, the decisive uncertainty, the evidence status and the hold or acceptance condition. The underlying heat map might remain in the annex as triage context. That arrangement respects the board's limited attention while exposing the main proposition. A score becomes a supplement to a decision record rather than a proxy for the decision itself.