THE GOLD STANDARD OF JUDGEMENT · SPECIALIST STUDY
Judge qualitative gains and losses around an explicit neutral level
Ordinary Sugeno aggregation ranks criteria on an ordered scale but does not by itself say whether an evaluation is favourable or unfavourable. Bipolar models add a neutral level, mirrored positive and negative levels and rules for aggregating gains and losses. That extra expressiveness also adds choices: symmetry, capacities, tie handling and cancellation must be declared and tested.
This guide belongs to the Judgement subject guide, the Gold Standard of Education, and the How X Works learning library. Choose a reading route, work the examples, then use the transfer checks to test whether the method survives a changed context.
Choose a reading route
Build the foundations · chapters 1–2
Apply the core method · chapters 3–5
Test evidence and alternatives · chapters 6–7
Handle limits and disagreement · chapters 8–9
Transfer and govern · chapters 10
The whole problem at a glance
| Element | Question | Risk if hidden |
|---|---|---|
| Neutral level | What counts as neither gain nor loss? | Arbitrary framing |
| Positive scale | How are favourable levels ordered? | False numerical interpretation |
| Negative scale | How are adverse levels mirrored or not? | Unjustified symmetry |
| Capacities | Which coalitions matter on each side? | Invisible interaction |
Ordinal levels support order and min/max operations; they do not automatically support distances, averages or subtraction.
CHAPTER 1 OF 10 · Build the foundations
1. Define a bipolar ordinal scale
Core idea. A bipolar scale orders adverse levels below a neutral element and favourable levels above it. In a symmetric model, negative levels mirror positive levels by order, not necessarily by measurable distance.
Method. Name every level, anchor the neutral reference and specify the sign-reversal operation. Elicit examples around boundaries and forbid arithmetic that the scale does not justify.
Worked example. A service is rated severe loss, mild loss, neutral, mild gain or strong gain. The labels order outcomes but do not say strong gain is twice mild gain.
Changed-context transfer. Transfer to learning support: below, at and above an expected reference can be ordered without pretending the gaps are equal.
Limits. People may disagree about the neutral reference and context can shift it. Record whose perspective and time horizon define neutrality.
Evidence and review. Convert this chapter into a dated claim register: record the evidence available, the assumption being made, the decision it affects, the person responsible, the strongest reasonable alternative and the event that would trigger reconsideration. Use the chapter’s practical check on real workflow output, not only on a demonstration case.
Misconception to avoid. Zero is naturally objective. Neutrality is a modelling and governance choice that can redistribute gains and losses.
Practical evidence check
- Ask three stakeholders to classify boundary cases and document every disagreement about neutral.
- Return to chapter contents
CHAPTER 2 OF 10 · Build the foundations
2. Separate positive and negative parts
Core idea. Bipolar evaluation decomposes a profile into favourable and adverse parts relative to neutral. This preserves the strongest reasons on both sides before a rule combines them.
Method. Transform each criterion into positive magnitude, negative magnitude or neutral, preserving the original label. Check that a criterion cannot be counted on both sides unless the model explicitly permits ambiguity.
Worked example. An option has strong safety gains and moderate privacy loss. Separate parts keep both visible rather than compressing them into an early net score.
Changed-context transfer. In medical choice, benefits and harms can be kept as ordered magnitudes before preference aggregation.
Limits. Decomposition can oversimplify mixed or ambiguous evidence. Use intervals or partial orders when classification is contested.
Evidence and review. Convert this chapter into a dated claim register: record the evidence available, the assumption being made, the decision it affects, the person responsible, the strongest reasonable alternative and the event that would trigger reconsideration. Use the chapter’s practical check on real workflow output, not only on a demonstration case.
Misconception to avoid. Every criterion is either good or bad. Neutral, unknown, mixed and inapplicable are different states.
Practical evidence check
- Classify ten cases and identify where a binary sign would erase uncertainty or context.
- Return to chapter contents
CHAPTER 3 OF 10 · Apply the core method
3. Use capacities for interacting criteria
Core idea. A capacity assigns ordered importance to coalitions of criteria, allowing complementarity and redundancy. In bipolar models, gains and losses may use one symmetric capacity or distinct positive and negative capacities.
Method. Elicit capacities from threshold cases and coalition comparisons, enforce monotonicity and retain rationale. Avoid asking stakeholders for every subset when a simpler structured family suffices.
Worked example. Safety and reliability together may be decisive even when neither alone is. A capacity for their coalition exceeds the levels assigned to individual criteria.
Changed-context transfer. In hiring policy, accessibility and training support may combine; ensure the coalition rule does not encode unlawful discrimination.
Limits. Capacity elicitation grows quickly and can hide inconsistency. Simplify, test and publish sensitivity rather than presenting parameters as discovered facts.
Evidence and review. Convert this chapter into a dated claim register: record the evidence available, the assumption being made, the decision it affects, the person responsible, the strongest reasonable alternative and the event that would trigger reconsideration. Use the chapter’s practical check on real workflow output, not only on a demonstration case.
Misconception to avoid. Capacities are just weights. They model coalition importance and need not add to one.
Practical evidence check
- Reconstruct each decisive coalition from an elicited comparison and check monotonicity.
- Return to chapter contents
CHAPTER 4 OF 10 · Apply the core method
4. Compute positive and negative Sugeno components
Core idea. Sugeno aggregation combines ordinal criterion levels with capacity thresholds using minimum and maximum. Applied separately, it identifies the strongest warranted positive level and strongest warranted negative level.
Method. For each distinct level, find criteria meeting or exceeding it, read the coalition capacity, take the minimum of level and capacity, then the maximum across thresholds. Repeat on the negative part.
Worked example. Two strong positive criteria form a high-capacity coalition, producing strong positive support; one moderate loss produces moderate negative support.
Changed-context transfer. In risk screening, repeat the logic for ordered protective and hazard indicators without turning labels into pseudo-probabilities.
Limits. Sugeno outputs can be insensitive to non-decisive changes and create ties. This is a feature of ordinal threshold aggregation, not a rounding error.
Evidence and review. Convert this chapter into a dated claim register: record the evidence available, the assumption being made, the decision it affects, the person responsible, the strongest reasonable alternative and the event that would trigger reconsideration. Use the chapter’s practical check on real workflow output, not only on a demonstration case.
Misconception to avoid. All criteria contribute smoothly. Max-min aggregation often depends on a decisive threshold and coalition; explanation should name them.
Practical evidence check
- Hand-compute both components and identify the exact threshold and coalition determining each.
- Return to chapter contents
CHAPTER 5 OF 10 · Apply the core method
5. Choose symmetric or asymmetric gain-loss treatment
Core idea. A symmetric model applies mirrored treatment to gains and losses, while cumulative-prospect-style Sugeno models can use separate capacities. Asymmetry may represent justified caution, but it can also encode unexamined bias.
Method. State whether positive and negative capacities differ, why, and for whom. Elicit both with comparable cases and test reversals when the framing changes.
Worked example. A safety decision gives severe losses stronger coalition importance than equally labelled benefits. The asymmetry is explicit and reviewed rather than hidden in wording.
Changed-context transfer. For education intervention, downside risks to vulnerable students may receive precautionary priority while benefits remain visible.
Limits. Loss aversion is not universally appropriate, and qualitative scales may not support numerical prospect-theory claims.
Evidence and review. Convert this chapter into a dated claim register: record the evidence available, the assumption being made, the decision it affects, the person responsible, the strongest reasonable alternative and the event that would trigger reconsideration. Use the chapter’s practical check on real workflow output, not only on a demonstration case.
Misconception to avoid. Asymmetry is irrational. It can be principled, but it must be transparent and proportionate.
Practical evidence check
- Swap gain and loss labels in matched cases and explain every change in the ranking.
- Return to chapter contents
Worked qualitative profile
| Criterion | Evaluation | Interpretation |
|---|---|---|
| Safety | strong positive | Above neutral and coalition-sensitive |
| Cost | moderate negative | Below neutral |
| Access | neutral | Neither gain nor loss under the declared reference |
| Reliability | strong positive | Can reinforce safety through capacity |
The final bipolar result depends on positive and negative capacities and the chosen combination rule; the profile alone does not dictate one answer.
CHAPTER 6 OF 10 · Test evidence and alternatives
6. Handle cancellation and symmetric maximum carefully
Core idea. Combining positive and negative components requires a qualitative analogue of subtraction. Symmetric maximum selects the sign of the stronger magnitude, but equal opposing magnitudes create a special cancellation case and associativity can fail.
Method. Declare the operator, tie rule and evaluation order. Keep component scores alongside the combined result so cancellation cannot hide serious reasons on both sides.
Worked example. Strong gain and strong loss cancel to neutral under one convention, yet the option is not reason-free. The report displays both components and marks a high-conflict neutral result.
Changed-context transfer. In policy review, a net neutral label can conceal concentrated harm and distributed benefit. Add distribution and rights checks outside the aggregator.
Limits. Different qualitative subtraction or lexicographic rules can resolve ties differently. No operator removes the need for judgement.
Evidence and review. Convert this chapter into a dated claim register: record the evidence available, the assumption being made, the decision it affects, the person responsible, the strongest reasonable alternative and the event that would trigger reconsideration. Use the chapter’s practical check on real workflow output, not only on a demonstration case.
Misconception to avoid. Neutral aggregate means nothing important is at stake. It may signal exact conflict between strong reasons.
Practical evidence check
- Create equal-magnitude gain-loss profiles and compare symmetric, lexicographic and component-wise outputs.
- Return to chapter contents
CHAPTER 7 OF 10 · Test evidence and alternatives
7. Interpret if-then selection and rejection rules
Core idea. Recent work relates bipolar Sugeno models to qualitative if-then rules. Threshold coalitions can be read as selection rules for favourable profiles and rejection rules for adverse profiles.
Method. Translate each decisive capacity threshold into plain-language conditions. Test rule coverage, overlap, exceptions and whether elimination rules are proportionate.
Worked example. If safety and reliability are at least high, select unless a severe rights loss triggers rejection. The rule exposes the threshold logic.
Changed-context transfer. In admissions, similar rules can become discriminatory; legal and fairness review must precede use.
Limits. Rule representations can be numerous and may not be unique. An explanation should identify the active rule, not claim the list is the only interpretation.
Evidence and review. Convert this chapter into a dated claim register: record the evidence available, the assumption being made, the decision it affects, the person responsible, the strongest reasonable alternative and the event that would trigger reconsideration. Use the chapter’s practical check on real workflow output, not only on a demonstration case.
Misconception to avoid. A readable rule automatically makes the decision fair. Transparency does not cure an unjust criterion or threshold.
Practical evidence check
- Generate active selection and rejection rules for three profiles and inspect disparate effects.
- Return to chapter contents
CHAPTER 8 OF 10 · Handle limits and disagreement
8. Diagnose ties, forgetfulness and weak discrimination
Core idea. Min-max operators are deliberately coarse and can ignore non-decisive details. Two alternatives may tie despite different secondary gains or losses; symmetric maximum can also forget the weaker side.
Method. Report positive and negative components, decisive thresholds and dominated reasons. Use lexicographic refinement, partial order or a richer model only when the decision need justifies it.
Worked example. Two options share the same strong positive result, but one has mild loss and the other moderate loss. A lexicographic refinement distinguishes them after preserving the primary ordinal result.
Changed-context transfer. In triage, a tie may appropriately trigger human review rather than a forced ranking.
Limits. Refinement can smuggle in unjustified distances or priorities. Preserve the original tie and explain the added rule.
Evidence and review. Convert this chapter into a dated claim register: record the evidence available, the assumption being made, the decision it affects, the person responsible, the strongest reasonable alternative and the event that would trigger reconsideration. Use the chapter’s practical check on real workflow output, not only on a demonstration case.
Misconception to avoid. More discrimination is always better. A coarse model can honestly represent limited preference information.
Practical evidence check
- For every tie, state whether it reflects true indifference, missing information or operator forgetfulness.
- Return to chapter contents
Sensitivity and tie matrix
| Change | Question | Required check |
|---|---|---|
| Move neutral level | Do signs change? | Recompute profile and ranking |
| Change one capacity | Does a coalition become decisive? | Elicit and compare rationale |
| Equal positive and negative magnitude | How is cancellation handled? | Declare tie convention |
| Refine ordinal scale | Do former ties split? | Test representation stability |
A stable ranking under one numeric coding is weak evidence when the intended model is ordinal.
CHAPTER 9 OF 10 · Handle limits and disagreement
9. Test neutrality, scale and capacity sensitivity
Core idea. Bipolar results can change when the neutral reference, ordinal granularity, capacities or tie convention changes. Sensitivity analysis should vary structural assumptions, not only numeric-looking parameters.
Method. Run plausible neutral levels, alternative scale refinements, monotone capacity sets and combination rules. Show stable, fragile and incomparable conclusions.
Worked example. Moving the neutral level reclassifies cost from mild loss to neutral and reverses two options. The report labels the ranking reference-dependent.
Changed-context transfer. Transfer to performance evaluation where “meets expectations” can shift with role or period.
Limits. The plausible set can be broad and elicitation costly. Use decision relevance to focus tests.
Evidence and review. Convert this chapter into a dated claim register: record the evidence available, the assumption being made, the decision it affects, the person responsible, the strongest reasonable alternative and the event that would trigger reconsideration. Use the chapter’s practical check on real workflow output, not only on a demonstration case.
Misconception to avoid. A single robust-looking output proves the model is objective. Robustness is always relative to tested alternatives and assumptions.
Practical evidence check
- Build a tornado-style assumption table without assigning fake distances to ordinal levels.
- Return to chapter contents
CHAPTER 10 OF 10 · Transfer and govern
10. Govern bipolar aggregation as decision support
Core idea. A bipolar integral structures judgement; it does not own the decision. Rights, distribution, uncertainty, evidence quality and accountability may impose constraints outside the ordinal aggregator.
Method. Publish scale, neutral level, capacities, operator, tie rule, sensitivity and reviewer. Preserve profiles and dissent, and provide an override with reasons and audit.
Worked example. A procurement ranking favours one option, but a mandatory accessibility condition excludes it. The record distinguishes model result from lawful decision.
Changed-context transfer. In classroom choice, use the model to discuss trade-offs rather than label a student or automate consequences.
Limits. Stakeholders can game criteria or neutral framing. Independent review and outcome monitoring remain necessary.
Evidence and review. Convert this chapter into a dated claim register: record the evidence available, the assumption being made, the decision it affects, the person responsible, the strongest reasonable alternative and the event that would trigger reconsideration. Use the chapter’s practical check on real workflow output, not only on a demonstration case.
Misconception to avoid. Mathematical form removes value judgement. The model makes some value choices visible but cannot eliminate them.
Practical evidence check
- Audit one decision from raw evidence through profile, aggregation, constraint, human reasoning and outcome.
- Return to chapter contents
Build a bipolar model card
Define purpose, alternatives, stakeholders, criteria and evidence before the scale.
Document neutral level, sign reversal, capacities, aggregation and tie handling with worked examples.
List prohibited uses, rights constraints and known blind spots.
Version the card and connect every decision to the exact model and evidence state.
Elicit capacities with threshold comparisons
Ask which coalitions are sufficient for each ordinal threshold instead of requesting abstract numbers.
Check monotonicity and resolve contradictions with participants rather than smoothing them silently.
Use sparse or structured capacity families when full elicitation exceeds available evidence.
Record disagreement and run alternative capacities through the decision.
Explain a bipolar result
Show the original criterion profile and neutral reference.
Name decisive positive and negative thresholds and coalitions.
Display both components before the combined result and reveal any cancellation or tie refinement.
State sensitivity, external constraints and who remains accountable.
Run a fairness and distribution audit
Test whether criteria, reference levels or capacities encode group disadvantage.
Compare burdens and benefits hidden by a net result.
Use counterfactual profiles and stakeholder review without treating protected traits as simple decision weights.
Monitor outcomes and provide remedy when the model contributes to unjust decisions.
Build a neutral-level elicitation workshop
Start with one live or realistic workflow and use concrete boundary cases to identify what stakeholders mean by neither gain nor loss. Define the decision, evidence owner, time boundary and affected people before selecting metrics. Preserve the original records so the exercise can be challenged and repeated rather than becoming a one-off narrative.
Work at least two contrasting cases through the same method. One should be ordinary and one should expose the boundary where a nominal zero can embed one group’s reference point as if it were universal. Record not only the final answer but also rejected interpretations, missing information and the consequence of delay or error.
Stress-test the result by changing one material assumption at a time. Ask which conclusion remains stable, which reverses and which becomes incomparable. Do not convert uncertainty into a hidden default merely to complete the table; route unresolved cases to proportionate review.
Close the exercise with an owner, correction route, retention rule and next review date. Publish the useful reasoning at the level permitted by privacy, security and law. Verification means that another authorised reviewer can reproduce the path and see exactly where judgement entered.
Create a capacity rationale ledger
Start with one live or realistic workflow and link every decisive coalition value to comparisons, evidence and a named decision perspective. Define the decision, evidence owner, time boundary and affected people before selecting metrics. Preserve the original records so the exercise can be challenged and repeated rather than becoming a one-off narrative.
Work at least two contrasting cases through the same method. One should be ordinary and one should expose the boundary where interaction parameters can become unexplained numbers that outlive the judgement that created them. Record not only the final answer but also rejected interpretations, missing information and the consequence of delay or error.
Stress-test the result by changing one material assumption at a time. Ask which conclusion remains stable, which reverses and which becomes incomparable. Do not convert uncertainty into a hidden default merely to complete the table; route unresolved cases to proportionate review.
Close the exercise with an owner, correction route, retention rule and next review date. Publish the useful reasoning at the level permitted by privacy, security and law. Verification means that another authorised reviewer can reproduce the path and see exactly where judgement entered.
Compare cancellation conventions
Start with one live or realistic workflow and run equal and near-equal positive and negative components through symmetric, lexicographic and component-wise rules. Define the decision, evidence owner, time boundary and affected people before selecting metrics. Preserve the original records so the exercise can be challenged and repeated rather than becoming a one-off narrative.
Work at least two contrasting cases through the same method. One should be ordinary and one should expose the boundary where a neutral headline can conceal strong conflict or depend entirely on a tie convention. Record not only the final answer but also rejected interpretations, missing information and the consequence of delay or error.
Stress-test the result by changing one material assumption at a time. Ask which conclusion remains stable, which reverses and which becomes incomparable. Do not convert uncertainty into a hidden default merely to complete the table; route unresolved cases to proportionate review.
Close the exercise with an owner, correction route, retention rule and next review date. Publish the useful reasoning at the level permitted by privacy, security and law. Verification means that another authorised reviewer can reproduce the path and see exactly where judgement entered.
Use partial orders when ranking overreaches
Start with one live or realistic workflow and retain incomparability when the model lacks enough justified preference information. Define the decision, evidence owner, time boundary and affected people before selecting metrics. Preserve the original records so the exercise can be challenged and repeated rather than becoming a one-off narrative.
Work at least two contrasting cases through the same method. One should be ordinary and one should expose the boundary where forced totals can create arbitrary winners and hide unresolved value conflict. Record not only the final answer but also rejected interpretations, missing information and the consequence of delay or error.
Stress-test the result by changing one material assumption at a time. Ask which conclusion remains stable, which reverses and which becomes incomparable. Do not convert uncertainty into a hidden default merely to complete the table; route unresolved cases to proportionate review.
Close the exercise with an owner, correction route, retention rule and next review date. Publish the useful reasoning at the level permitted by privacy, security and law. Verification means that another authorised reviewer can reproduce the path and see exactly where judgement entered.
Backtest qualitative thresholds
Start with one live or realistic workflow and compare past profiles, active rules, outcomes and stakeholder review without converting ordinal levels into false measurements. Define the decision, evidence owner, time boundary and affected people before selecting metrics. Preserve the original records so the exercise can be challenged and repeated rather than becoming a one-off narrative.
Work at least two contrasting cases through the same method. One should be ordinary and one should expose the boundary where a rule may appear coherent in examples yet fail on rare coalitions or changed contexts. Record not only the final answer but also rejected interpretations, missing information and the consequence of delay or error.
Stress-test the result by changing one material assumption at a time. Ask which conclusion remains stable, which reverses and which becomes incomparable. Do not convert uncertainty into a hidden default merely to complete the table; route unresolved cases to proportionate review.
Close the exercise with an owner, correction route, retention rule and next review date. Publish the useful reasoning at the level permitted by privacy, security and law. Verification means that another authorised reviewer can reproduce the path and see exactly where judgement entered.
Questions readers ask
What makes a Sugeno model bipolar?
- It adds a neutral level and ordered positive and negative evaluations, then aggregates gains and losses under declared rules.
Are positive and negative distances numerical?
- Not necessarily. The model can be purely ordinal.
Must the model be symmetric?
- No. Separate positive and negative capacities can represent justified asymmetry.
What does a neutral result mean?
- It may mean neutral evidence or cancellation between equally strong gains and losses; show the components.
Why are there many ties?
- Max-min aggregation is coarse and threshold-based.
Can the result decide automatically?
- It should support accountable judgement and remain subject to rights, law and governance.
Practise, verify and continue
Create a five-level bipolar scale and three alternatives. Elicit positive and negative coalition thresholds, compute both Sugeno components, apply two tie rules and test a shifted neutral level. Explain every ranking change without treating levels as equal intervals.
Transfer the model to a school, health or procurement example. Add a rights constraint and show where aggregation ends and accountable judgement begins.
Related eduKate reading
Sugeno integrals, ordinal capacities and non-compensatory choice
Choquet integrals, capacities and decision attitudes
Possibility theory, necessity measures and fuzzy evidence
Stochastic dominance and partial preferences
Authoritative sources and scope
Dubois and co-authors, Bipolar Sugeno integrals, if–then rules and preference orderings
Michel Grabisch, The Symmetric Sugeno Integral
Grabisch, Labreuche and Vansnick, aggregation of interacting criteria and neutral levels
Grabisch and Labreuche, bipolarization of ordered structures
This article is general education in ordinal decision modelling. It does not recommend automated high-stakes decisions or imply that qualitative labels are measurements. Models depend on evidence, neutral reference, scale, capacities and tie rules. Preserve components, disagreement and sensitivity; apply legal, rights and distribution checks outside the aggregator. Examples are original teaching fixtures. The gold standard uses the model to expose reasons and limits, not to turn contested values into an unquestionable score.