Cappelen-Plunkett Conceptual Engineering

Short answer: yes, re-engineer — but the description you’ve already done is what licenses the re-engineering, so it’s not “rather than.” The four-jobs diagnosis is the engineering input. And note a wrinkle in citing Cappelen and Plunkett together, because they don’t pull in quite the same direction, and the difference matters for how you implement this.

The method, and the C&P wrinkle

The function-first move you’re reaching for is really Haslanger’s ameliorative question — not “how is ‘fairness’ used?” but “what concept should this term express, given the work we need it to do?” — which the Cappelen/Plunkett/Burgess program absorbs as a central case of conceptual ethics. Plunkett (with Sundell) adds the useful reframe that your four-way overload is a set of covert metalinguistic negotiations: when two managers “disagree about whether the policy is fair,” they’re often not disagreeing about facts under a shared concept — they’re tacitly advocating different concepts while using the same word. That’s the disease. Engineering is the cure precisely because the dispute is really about which concept ought to govern.

The Cappelen wrinkle cuts the other way and you should take it seriously: in Fixing Language he’s skeptical that we can control what our words come to mean — meaning is anchored by use across a whole community, not by a definition you legislate in a policy doc. The practical upshot isn’t “don’t engineer.” It’s: don’t expect a stipulated definition of ‘fairness’ to hold. The word is too overloaded; people will keep pouring their preferred job back into it. That argues for a sharper move than redefinition — see implementation below.

Sharpening the diagnosis: which jobs actually conflict

‘Fairness’ here isn’t vague (one fuzzy meaning) — it’s overloaded (four distinct meanings). The conflicts aren’t symmetric:

• Equal treatment (same process for all; “anti-classification”) vs equal outcomes (representation tracks the qualified pool; “anti-subordination”) — the deep, irreducible conflict. A blind process can yield skewed outcomes; equalizing outcomes requires treating people differently by group.
• Meritocracy vs equal outcomes — if the qualified pool is itself unequal, selecting the best-qualified produces unequal outcomes. Same conflict, different surface.
• Bias-correction is the interesting one: it looks like a fourth rival, but it can be folded into meritocracy without contradiction — if you frame it as correcting the signal, not lowering the bar. Discounting that a candidate couldn’t afford an unpaid internship isn’t overriding merit; it’s removing a known distortion in the inference from signal → actual ability. That’s more accurate meritocracy, not less.

So the four collapse to a stable core plus one genuine outlier. Three of them (equal-treatment-of-process, bias-corrected assessment, meritocratic relevance) compose into one coherent concept. Equal-outcomes is the one that cannot be folded in, because it’s a population-level property and a hiring decision is a per-candidate act — you literally cannot evaluate whether a single hire was “fair” in the equal-outcomes sense without incoherence.

What ‘fairness’ should mean for this policy

My recommendation: engineer ‘fairness’ to name the per-candidate evaluation standard, and explicitly carve representation out of it as a separate, separately-named objective. Concretely, the operative definition:

A hiring decision is fair when each candidate is assessed against criteria demonstrably relevant to job performance, using the same process for every candidate, with deliberate correction for known distortions in the signals we rely on (credential access, interview artifacts, biased proxies). Fairness is a property of how we evaluate individuals. Workforce representation is a distinct organizational objective, pursued through [named levers: sourcing, pipeline, criteria design], not part of the fairness standard.

Why this target:

• It’s coherent — the four-way overload resolves to one concept once bias-correction is read as merit-accuracy and outcomes are carved off.
• It’s action-guiding at the point of decision — a hiring manager can actually apply it to the candidate in front of them, which “did this hire produce equal outcomes?” can never be.
• It honors the function — hiring is per-candidate selection; the concept that does work at the moment of choosing is “assess this person accurately and on relevant grounds.”
• It makes the real conflict negotiable instead of covert — representation doesn’t disappear; it gets pursued under its own name, with its own justification, traded against fairness explicitly rather than smuggled in through a contested word.

Where the value commitment sits — be honest about this

This engineering choice is not neutral, and you should own that rather than present it as the analysis “discovering” the answer. Folding bias-correction into “merit accuracy” still rests on a meritocratic core — and an anti-subordination theorist will say “job-relevant criteria” and “merit” are themselves shaped by prior advantage, so my reconciliation launders structural injustice one level up (you’ve debiased the signal but not interrogated whether the bar encodes privilege). That objection is live and I don’t think it’s defeated; it’s located. The load-bearing decision is: does your policy treat hiring as primarily a per-candidate desert question, or primarily a lever for distributive justice? My proposal answers “the former, with the latter pursued separately and transparently.” If your organization’s actual commitment is the reverse, you’d engineer ‘fairness’ toward the substantive/anti-subordination meaning and carve out procedure instead. Either is defensible; what isn’t defensible is leaving both inside one word.

Implementation — and the Cappelen-driven sharper move

Given that you can’t fully control the word: consider retiring “fairness” as the load-bearing term in the policy text entirely. Replace it with the named sub-concepts — relevant-criteria assessment, equal process, signal de-biasing, representation — and reserve “fair” for informal summary. This is the cleaner conceptual-engineering outcome when a term is this overloaded: lexical retirement beats redefinition, because a redefinition just gives everyone a new surface to project their preferred job onto. Then add one explicit clause naming how the fairness standard and the representation objective get traded off when they conflict (e.g., representation operates on sourcing and criteria-design upstream, the fairness standard governs the decision — so they rarely collide at the same stage by construction). That’s the design that makes the genuine equal-treatment/equal-outcomes conflict manageable instead of perpetually re-litigated under one ambiguous word.

If it’d help, I can draft the full policy-section language with the four concepts named and the trade-off clause spelled out — say the word and I’ll write it.

Yes, following the methodological mandates of Conceptual Engineering (specifically the Cappelen-Plunkett approach to meaning and utility), you should absolutely re-engineer the concept of “fairness” in this context.

Here is the breakdown of why you should not simply describe the current usage, what specific engineering is required, and what definition you should adopt.

1. The Diagnosis: Why Description Fails

Description accepts the world as it is used by speakers. If you merely describe “fairness” in your hiring policy, you accept that the term is overloaded. This leads to conceptual confusion, where:

• Hiring managers think “fairness” means equality of opportunity (merit).
• HR thinks “fairness” means peace (bias correction).
• Employees think “fairness” means protection (equal outcomes/equity).

As Daniel Cappelen argues in his framework (e.g., Why there shouldn’t be a theory of concepts), much of philosophical puzzlement arises not because language is vague, but because the words are doing too much. The inherited concept “fairness” is technically a settled concept in ordinary language, but a failed concept in policy execution. It does not map cleanly to legal eligibility or functional policy mechanics.

2. The Engineering Mandate: Fix the Tool

Plunkett (in the context of Conceptual Engineering applied to law/social policy) agrees that if a concept creates friction or prevents specific judgments from being made clearly, we must intervene.

The goal is not to find “God’s definition” (descriptive) but to find the definition that serves the policy intent.

The New Definition of ‘Fairness’ for the Policy: In a hiring context, “fairness” should be re-engineered to mean Procedural Neutrality + Non-Discrimination.

Operational Definition: Fairness in this hiring policy shall be defined as the application of uniform, non-discriminatory selection criteria based on job-relevant competencies, where bias-correction measures are transparently applied as sub-components of that process, not the end goal.

Why this definition works:

1. Severs the “Outcome” Trap: It explicitly removes “equal outcomes” from the definition. You are not promising the result is equal; you are promising the process is impartial.
2. Bindings “Merit” and “Bias” together: It defines merit not as “natural superiority” but as “competence relative to the role,” while bias-correction is a mechanism to ensure the assessment of that merit is clean, not a substitute for merit.
3. Resolves the Incompatibility: It makes the policy honest about the tension. You aren’t claiming to achieve both equal treatment and equal betterment; you are claiming to achieve Equal Access.

3. Implementation: Splitting the Lexicon

Cappelen and Plunkett would likely recommend that “Fair” be split into more granular terms to prevent the “concept drift” you are currently suffering from.

Proposed Polysemous Engineering:

4. The Rationale (The Cappelen Argument)

If you describe “fairness” without engineering, you invite conceptual friction.

• The Problem: If a manager rejects a candidate to “maintain fairness,” they often mean “merit” (the concept is expensive). If an employee complains, they mean “outcomes” (concept is expensive).
• The Solution: Acknowledge that “fairness” is a metaphor for two distinct things:
  1. The Law: Anti-discrimination (Prohibits negative impact).
  2. The Ethics: Moral desert (Allows for positive impact/merit).
• The Fix: The policy must say: “Fair treatment is defined by compliance with A and B. We are not guarantee C.”

Summary Recommendation

You should move the policy from Descriptive Fairness (what people think “fairness” is) to Normative Engineering (what “fairness” should do to function).

What it should mean: Fairness should be redefined as “Selection Integrity”.

• Integrity signals transparency about the process.
• It implies that if you follow the rules, you are treated fairly, regardless of the outcome.
• It separates “bias correction” (removing the poison) from “meritocracy” (choosing the best wine).

By re-engineering it this way, you stop using one word to try to heal two disjointed parts of your policy (legal compliance vs. moral perfection) and instead use precise language to ensure stakeholders understand which function the policy is fulfilling in any given moment.

Target concept

Target concept: Fairness (as operationalized in organizational hiring policy).

Engineering question: If four functions (equal treatment, equal outcomes, bias-correction, meritocracy) are conceptually conflated in current usage, can ‘fairness’ be re-engineered to resolve these functional conflicts?

Assumption: This analysis proceeds on prompt-derived assumption that functions ARE conflated; without policy text access, cannot verify baseline claim with confidence.

Current usage — descriptive baseline

Baseline 1: Functions used: Equal treatment (assessing candidates identically regardless of demographic), Equal outcomes (demographic parity at hire date), Bias-correction (systemic discrimination actively countered), Meritocracy (best candidates selected regardless of process).

Baseline 2: Functions often conflated. Equal treatment conflicts with outcome-based justice. Very common theoretical conflict.

Baseline 3: Prompt-specified four functions exist as conceptual baseline, per prompt wording. Policy text not accessed; functions treated as analysis assumptions, not empirically verified facts.

Baseline 4: Function conflations are not strictly policy-text-specific; logical tensions exist at the level of hiring decision theory.

Baseline 5: Conceptual baseline states current usage conflates these functions; prompt-derived assumption, not empirically verified without policy/process documents.

Identified function failures

• Failure 1 (Legitimation): Fairness should justify hiring decisions to stakeholders; current failure allows concurrent claim conflict (Merit-stakeholders vs. Outcomes-stakeholders) with no definitive meaning.
  • Cost of Failure 1: Stakeholders argue past each other; no single decision is definitive.
• Failure 2 (Coordination): Fairness should specify concrete intervention points; current failure creates decision ambiguity when manager rejects candidate to meet outcome target (no concept guidance on what ‘fair’ means in this context).
  • Cost of Failure 2: Managers hesitate; document review slows; reviews note ‘fairness concerns’ that cannot be resolved.
• Failure 3 (Moral guidance): Fairness claims direct action toward values; current failure creates value overload where concept frames the goal as ‘Fairness’ (singular) rather than ‘Just X’.
  • Cost of Failure 3: Organization speaks of ‘living values’ while people cannot operationalize them.
• Failure 4 (Justificatory): Fairness claims should be defensible evidence-based claims; current failure allows ‘We’re fair’ justified by any function depending on defender.
  • Cost of Failure 4: Moral hazard, reform barriers.
• Failure 5 (Action-guiding): Fairness should enable hiring managers to make decisions; current failure creates decision paralysis, policy whiplash, credibility loss.
  • Cost of Failure 5: Decision paralysis, policy whiplash, credibility loss.

Cost of All Failures: Generalizable to organizational context with these functions; depends on hiring policy specifics.

Root Cause: Concept conflates decision logic (how we select individuals) with impact audit (what we achieve); cognitively merged into one term.

Ameliorative purpose

Ameliorative-purpose function: Fairness in hiring policy should be able to explain why a specific hiring decision satisfies organizational commitments without requiring subsequent justification to reconcile conflicting sub-claims.

Constraint: This function accommodates organizational substantive commitments without carrying the compromise artifact of ‘fairness-now-means-unknown.’

Alternative purpose phrasing: Fairness in hiring policy should clarify which function governs which decision point by design, not word meaning.

Note: Function-shape stated (something the revised concept should do), not conclusion-shape (desired classification the concept should produce).

Ameliorative-Overreach Note

Note: this concept’s contestation may be constitutive (essentially contested in Gallie’s sense); the engineering move risks treating as resolvable what is in fact the object of legitimate ongoing dispute. Candidate revisions below are offered with this caveat.

Overreach risk: Treating fairness as something that SHOULD mean one specific thing risks overreach when term’s contested status is constitutive.

Assessment: Fairness in hiring IS essentially contested (parties have good counter-arguments depending on function invoked). If so, engineering one definition as ‘the right one’ exceeds framework’s legitimate scope.

Alternative move: Engineer the function, not the definition — clarify which function governs which decision point by design, not word meaning.

Flag Status: Applicable; emphasis on functional disambiguation rather than definition-as-doctrine.

Candidate revisions

Candidate 1 (Process Fairness/Procedural Fairness Only)

Claim: Fairness = adherence to predefined, role-specific process that includes bias-correction and outcome-monitoring but is NOT defined by outcome targets OR Fairness = Consistent application of selection criteria irrespective of candidate background or resulting distribution.

Rationale: Separates process from outcome; retains bias-correction as process requirement OR Aligns with legal ‘integrity’ doctrines; eliminates the conflict by removing outcomes/bias-correction from the concept itself.

Tradeoff: Process gaming risk; doesn’t ensure diversity outcomes OR May reduce ability to justify bias-correction measures as ‘fair’.

Candidate 2 (Commitment Fairness)

Claim: Fairness = specific impartiality commitment for relevant job classification at time t, explicitly documented and signed off by hiring authority.

Alternative Claim: Fairness = documents the subset of hiring practices a decision-maker has bound themself to, according to explicit definition for that position.

Rationale: Concept becomes operational document rather than contested normative claim OR Documentations instruments become the standard, not the concept; shifts ethical weight to documentation.

Tradeoff: Administrative overhead; shifts ethical weight to documentation OR Loss of unified fairness symbol as organizational language.

Candidate 3 (Integrity + Stewardship / Biconceptual Split / Contextualized Stewardship)

Claim: Fairness splits into Selection Integrity (procedure + merit, no outcomes weighting) and Equity Stewardship (outcomes + bias-correction, separate audit metric) OR Fairness contextually labeled: Fair procedure during interview/assessment, Fair outcome during representation tracking, Fair representation during impact assessment.

Rationale: Resolves incompatibility by removing ‘outcomes’ from the decision-criteria definition; distinguishes how we hire from who we hire OR Embraces the context-dependency rather than trying to define a universal fairness meaning.

Tradeoff: ‘Fairness’ becomes a compound system; moral shorthand value partially lost; Fairness becomes an audit metric OR More terminological burden; may feel evasive to stakeholders.

Recommendation Note: Candidate 1 split is strongest engineering move, preserves ‘Fairness’ term for decision while creating ‘Stewardship’ for audit; Candidate 2 Commitment Fairness preferred ONLY if organization chooses operational accountability over symbolic coherence.

Note: Neither analysis makes universal recommendation; ties to organizational values decision.

Implementation problem

Problem Stage 1 (CAPM conceptual uptake): External conceptual engineering literature signals implementation depends on community uptake (Cappelen & Plunkett 2019); precise mechanism remains ambiguous.

Problem Stage 2 (Organizational decision variables): Legal risk exposure; stakeholder veto power; contractual constraints; cultural readiness metrics; resource availability.

Problem Stage 2 (Specific): Rewire definitions, KPIs, policy language; Accept explicit vs implicit commitments; Verify defensible against legal standards; Accommodate new fairness signals in ATS; Understand fairness instrument meaning.

Two-Stage Structure: Modeled as: Stage 1 = Community/cultural notion-shifting; Stage 2 = Organizational decision-to-implementation variables.

Implementation Risk: High (without verified policy text, adoption risk cannot be quantified).

Adoption Mechanism: Policy redraft; Training scripts; Legal review; Communication (‘clarify language’ not ‘change values’); Timeline (3-6 months to shift cultural shorthand).

Stakeholder Parties: HR leadership, hiring managers, Legal/Compliance, HR tech, external stakeholders.

Revision costs and displacement

• Cost 1 (Moral shorthand): ‘We are a fair company’ relies on vague conflation; re-engineering requires ‘We are an integrity-focused company with explicit equity stewardship.’
• Cost 2 (Psychological assurance): Hiring managers rely on ‘Fairness’ to mean ‘Just hire the best’; Commitment Fairness ‘implies moral burden’ they want to avoid; Selection Integrity/Integrity + Stewardship adds cognitive load.
• Cost 3 (Linguistic/terminological learning): Staff must learn new terminology; ATS/HR tech customization required.
• Cost 4 (Legal precedent): ‘Integrity’, ‘Selection Integrity’, or ‘Commitment Instrument’ has no established legal precedent; Counsel may require explicit testing; ‘Documented commitment unless defined’ may be interpreted as discriminatory.
• Cost 5 (Methodological credentialing): ‘Loss of methodological consensus language lacks external validation credentialing’; Revision cannot cite ‘Cappelen-Plunkett recommended this’ with confidence.
• Cost 6 (Symbolic vs operational tradeoff): All revisions displace more symbolic comfort for more operational control; Which tradeoff is worth exchanges depends on organizational values choice.
• Cost 7 (Contractual constraints): Existing agreements potentially reference ‘fair’ as outcome term; Employment agreements may bind party to ‘fair treatment’ language.

Confidence per finding

• Function-Failure Confidence: High (logical incoherence of Equal Treatment vs Bias Correction is known structural problem in social management theory, independent of policy text); Medium (depends on actual hiring context; depends on policy transparency); Medium (generalizable to organizational context); High (on analytical coherence given stated methodological constraints).
• Proposed-Revision Confidence: High (on analytical coherence given stated methodological constraints); Medium (on adoption risk given policy text absence); Low (on formal methodological verification per Cappelen-Plunkett standards); Low (CAPM procedure phases, Phase 5 engineering moves, Phase 6 adoption).
• Adoption Feasibility Confidence: Low (CAPM procedural requirement ambiguous); Low-Medium (Implementation problem significant and likely outside immediate proposal control); Medium (generalizable to organizational context); Low (organizational decision variables not empirically mapped); Low (CAPM implementation depends on community uptake, mechanism ambiguous).
• Baseline Confidence: Medium-High (on prompt statement); Low (on policy text verified); Medium (depends on actual policy wording).

Additional considerations

Gaps & Uncertainties

• Gap 1 (CAPM framework specifics): Direct access to Cappelen & Plunkett published methodology for ‘ameliorative analysis’ and ‘conceptual containment’ not accessed; RAG returned null.
• Gap 2 (Policy text access): Four claimed functions (equal treatment, outcomes, bias-correction, meritocracy) treated as assumptions from prompt rather than verified from actual policy document.
• Gap 3 (Legal specifics): Legal risk for ‘commitment-based fairness instruments’ not mapped against specific EEOC/employment law frameworks.
• Gap 4 (Cultural readiness metrics): Assumed from general management sources; not organization-specific.
• Gap 5 (Methodological verification): Both RAG attempts returned null; output’s Cappelen-Plunkett framing asserted, not verified. Analysis proceeds on analytic coherence alone.

Sources & Evidence Attribution

• CAPM framework sources consulted: philpapers.org/plunkett, plato.edu archives, burger 2019, seyfriends engineering texts.
• Organizational management sources consulted: mindfulexecutivestrategies.com, mckinsey.com, 2030.builders, indeeed.com, beehivepr.biz.
• Verification status: Unclear whether external sources explicitly address ‘Fairness’ fairness-policy mapping; Specific ‘commitment documentation’ approach not directly sourced from published framework text.
• Methodological caveat: ‘As Cappelen-Plunkett methodology suggests but cannot confirm from available sources’ — hedged per available evidence.

Claim-Resolution Status

• Claim 1 (Community uptake requirement): Ambiguous / Disputed. External conceptual engineering literature signals implementation depends on community uptake (Cappelen & Plunkett 2019); precise mechanism remains ambiguous.
• Claim 2 (Symbolic vs operational tradeoff): Disputed. Revision cost tradeoffs hypothesized for organizational contexts; claim about ‘symbolic comfort’ displacement lacks specific empirical source; retained with contested framing.
• Claim 3 (Framing implementation): Disputed. Specific fairness-policy mapping proposed here not directly sourced from published framework text; retained with contested framing.

Target concept

“Fairness” as operationalized within the organization’s hiring policy. The engineering question on the table is: given that the concept is overloaded with four incompatible functions (equal treatment, equal outcomes, bias-correction, meritocracy), should it be re-engineered through prescriptive or normative revision rather than left descriptive of current usage, and if so, what should it mean?

Current usage — descriptive baseline

• Equal treatment — same process, same criteria, applied uniformly; procedural blindness to protected characteristics at evaluation. Implicit theory: fairness is blindness (the rule does not see the person).
• Equal outcomes — demographic parity or proportional representation in hires. Implicit theory: fairness is balance (the result mirrors the pool).
• Bias-correction — active counterweight to historical/systemic disadvantage in pipeline or evaluation. Implicit theory: fairness is rectification (equity restores what was taken).
• Meritocracy — selection on validated, job-relevant competencies / highest predicted performance. Implicit theory: fairness is desert (the capable earn the position).

The policy currently treats “fairness” as a monolithic concept expected to deliver these four distinct jobs simultaneously. The four implicit theories carry contradictory inferential commitments: blindness treats group membership as irrelevant while mirroring treats it as essential; tournament/desert treats the input pool as legitimate while rectification treats it as contaminated. When a hiring manager invokes “fairness” they reach for one source while the policy name covers all four, which is why one word authorizes opposite decisions.

It should be noted that the four-function list may not be exhaustive—a fifth function (dignity, opportunity) may be silently present and worth surfacing. Furthermore, the four functions are not logically incompatible in all combinations, only in their naive, strong forms. Partial reconciliation is available if “equal treatment” means “apply the corrected rules equally,” “merit” is defined post-correction, and “equal outcomes” targets the relevant talent pool rather than 50/50. The descriptive baseline parses “equal treatment” as procedural blindness, but the precise statutory definition of equal treatment varies by jurisdiction (e.g., US EEOC vs. UK Equality Act 2010), so final policy drafting requires alignment with local employment law.

Identified function failures

• Equal treatment vs. bias-correction — current concept’s failure mode: Strict procedural blindness prevents the active, targeted interventions that bias-correction requires. The functions are mutually antagonistic, not merely distinct. Cost of the failure: operational paralysis.
• Meritocracy vs. equal outcomes — current concept’s failure mode: Maximizing individual predicted performance will, under unequal starting conditions or biased upstream pipelines, systematically conflict with achieving aggregate demographic parity. Cost of the failure: contradictory outcomes authorized under one banner.
• Decision rule for clashes — current concept’s failure mode: Because one word covers all four functions, the policy supplies no rule for when they conflict. Cost of the failure: Hiring managers face operational paralysis or act arbitrarily; the organization invites legal and reputational risk by defending contradictory outcomes under one banner. The concept functions as a rhetorical shield rather than an operational guide.
• Descriptive mode capability — current concept’s failure mode: Reporting current usage would only confirm the incoherence; only the engineering mode (prescribing what the concept should mean) is responsive to the diagnosed problem. Cost of the failure: a concept asked to denote four things whose strong forms are mutually exclusive cannot reliably denote any of them. A concept is defective when it fails to track what you want it to track; the response is to redesign the tool, not the task.

Ameliorative purpose

The revised concept should function as a transparent conflict-resolution and priority-ranking mechanism, guaranteeing that when legitimate organizational values (merit, bias-correction, representation) inevitably conflict, deviation from strict proportional merit is governed by pre-declared, auditable rules rather than hidden ad-hoc judgment. The choice of which function(s) the policy is in the business of serving is itself an engineering decision and should be made explicitly.

The sub-functions the revised concept should support are:

• Explicitly prioritizing organizational goods (e.g., a default of talent acquisition, workforce representativeness, and procedural integrity, with rectification as a constraint on how the first three are operationalized rather than as a fourth independent goal).
• Replacing rhetorical ambiguity with auditable, determinate rules without smuggling a specific conclusion (like “demographic parity” or “strict colorblindness”) into the purpose statement itself.

Note: this concept’s contestation may be constitutive (essentially contested in Gallie’s sense); the engineering move risks treating as resolvable what is in fact the object of legitimate ongoing dispute. “Fairness” in hiring is an essentially contested concept: the dispute about what it means is partly constitutive of the underlying political disagreement about how hiring should work. No engineering effort can produce a definition that achieves universal philosophical consensus or permanently resolves the merit-vs-equality tension. The engineering move proposed is therefore an ameliorative project (a proposal that the concept should mean this), not a discovery (that it does mean this). The policy is being made more determinate, not aligned with what fairness “already is.” This caveat is reframed structurally: the definition is defeasible, holding absent further contestation and revisable through a stated procedure. If the value hierarchy shifts, the re-engineering is itself a candidate for re-amendment, meaning an amendment mechanism (e.g., biennial review of the construct/weighting matrix by a cross-functional governance board) should be built into the policy.

Candidate revisions

Candidate 1: Proceduralism. Fairness = transparent, consistent, job-relevant criteria applied uniformly. Rationale: simplest, most legally defensible, treats candidates as individuals. Tradeoffs and costs: explicitly disavows bias-correction and equal outcomes — gives up on representation and rectification; if the input pool or criteria themselves carry historical bias, uniform application ratifies that bias while appearing neutral.

Candidate 2: Auditable Procedural Equity (upstream/downstream split). Fairness is achieved when the process separates pipeline intervention from candidate evaluation: bias-correction applied upstream (anonymized screening, diverse slates); final evaluation strictly meritocratic on pre-defined job-relevant criteria; equal outcomes treated as a diagnostic metric triggering upstream process reviews, not as a target of individual hiring decisions. Rationale: cleanly separates incompatible functions — satisfies bias-correction without violating equal treatment at the point of evaluation, protects meritocracy while addressing systemic issues. Tradeoffs and costs: demotes “equal outcomes” from a definition of fairness to a diagnostic trigger; stakeholders who view demographic parity as the only true measure of fairness find it inadequate; relocates the conflict upstream without resolving residual trade-offs; applies a rigid, role-agnostic split.

Candidate 3: Equity-primacy. Fairness = active correction for documented group-level disadvantage, with procedural and merit components subordinated to that end. Rationale: addresses root cause (historical/structural inequity) rather than symptom (biased outcomes). Tradeoffs and costs: abandons strict equal treatment and pure meritocracy; requires ongoing, empirically and politically contested documentation of disadvantage; if the rectification target is mis-specified, corrects the wrong thing while appearing rigorous; legally vulnerable in strict anti-preference jurisdictions.

Candidate 4: Transparent Value-Weighting (role-specific). Fairness = the organization’s commitment to openly state the context-dependent weighting of merit, bias-correction, and representation for each specific role, and to apply that stated weighting consistently (e.g., specialized technical role weighted merit 90% / bias-correction 10%; leadership role weights representation higher to correct historical homogeneity, provided minimum merit thresholds are met). Rationale: treats all four functions as valid organizational goods, refuses to let them hide inside a single ambiguous word, makes trade-offs explicit and auditable, scales to role-specific variation. Tradeoffs and costs: significant administrative overhead to define and justify per-role weightings; exposes the organization to criticism for explicitly acknowledging that merit is sometimes subordinated to other goals.

Candidate 5: Coordinated Multi-Component Construct (priority-ordered decomposition). Fairness is not a single property but a family of four explicit, named requirements, each addressing one function, with a stated priority order for conflict cases:

1. Process integrity — transparent, consistent, job-relevant criteria; same procedure applied to all candidates at the same stage; appeal path for procedural defects. (Hard floor — no outcome justifies an unfair process.)
2. Bias-correction — documented, time-bounded adjustments to criteria or process where evidence shows the unadjusted procedure produces systematically adverse impact. (“Documented” and “time-bounded” are load-bearing; without them bias-correction is unfalsifiable. Operates on criteria, not individual decisions.)
3. Representational monitoring — periodic measurement of hire outcomes against the relevant labor-market availability pool (not 50/50, not internal incumbency), with corrective review on statistically significant shortfalls. (Monitoring + review, not a hard quota; triggers review but does not override individual decisions.)
4. Merit assessment — selection on validated job-relevant competencies, validated against subsequent job performance where measurable. (Tiebreaker among candidates who clear the first three.) Rationale: honors the diagnosis directly — the four functions are not the same thing, so a unified concept cannot serve them; gives the word determinate content; surfaces disputes as disputes about which component applies rather than what fairness means; builds in a falsifiability condition; preserves all four functions rather than collapsing to one. Operational thresholds make this auditable: adverse-impact trigger operationalized as the 4/5ths (80%) rule; relevant labor-market pool defined as a published weighted composite of recruitment geography, feeder-pipeline qualification-tier, and published occupational workforce data; measurement cadence stated and results disclosed where legally permissible. Tradeoffs and costs: loses conceptual unity (the diagnostic shows that unity was illusory); longer and more technical. Note: the claim that bias-correction “operates on criteria, not individual decisions” holds cleanly in some cases but bleeds in others. When criterion-level and case-level adjustments cannot separate cleanly, the priority order is explicitly invoked, the finding documented, and the decision routed through escalation rather than absorbed silently.

Candidate 6: Hybrid — Auditable Procedural Equity with Role-Weighted Residual Tradeoffs. Rationale: explicitly combines upstream bias-correction (diverse slates, anonymized screening) with role-specific weighting for final evaluation. It addresses the upstream-ambiguity costs that role-weighting alone leaves on the table while retaining auditable per-role conflict-resolution. Tradeoffs and costs: aggregates the administrative burdens of both Candidate 2 and Candidate 4, though it resolves the upstream ambiguity.

Surfaced tension: The two recommended anchors — role-specific value-weighting (Candidate 4) and priority-ordered component decomposition (Candidate 5) — are different engineering moves serving the same ameliorative purpose. One makes the weighting between functions the explicit, per-role, auditable object; the other makes the named components and their fixed priority order the explicit object, with representation demoted to monitoring rather than a weighted factor. Both replace rhetorical ambiguity with auditable rules; they differ on whether equal-outcome considerations enter as a weighted input (Candidate 4) or as a monitoring trigger (Candidate 5). The organization, rather than the analysis, selects between them.

Implementation problem

Proposing a revision is not the same as achieving adoption. The current ambiguity is politically useful precisely because factions project their preferred definitions onto it. However, the problem is more tractable here than the general implementation worry because policy documents are formal, definition-bearing artifacts, the relevant population is bounded (hiring managers, recruiters, legal, candidates via published policy), and the institutional lever is direct. The organization can adopt a revised definition by writing it, training on it, and auditing against it.

Adoption requires concretely: rewritten policy text detailing the four components or weightings, priority order, operational thresholds, pool-scoping methodology, and legal clauses; decision-support tools like rubrics and scorecards embedding the 4/5ths rule into ATS workflows; training for hiring committees to recognize which component governs each case and when criterion-level and case-level bias-correction have bled together; and candidate-facing communication publishing what “fairness” means and its methodology. A top-down executive mandate is needed to override comfortable ambiguity; without enforceable auditing, actors will revert to using “fairness” as a catch-all.

Stakeholder mapping highlights specific adoption conditions: strict meritocrats will require merit to remain the default baseline weighting for most roles with high evidentiary burdens for deviations; diversity and equity advocates will require role-specific weightings to robustly address historical underrepresentation backed by transparent public auditing; and legal/compliance officers will require the weighting matrix to be pre-vetted for statutory compliance, offering a stronger, more defensible legal record than ad-hoc ambiguity.

A residual gap remains: even after full adoption, candidates, rejected applicants, and external stakeholders will read “fairness” through their own inherited concept, and disputes will continue. Where stakeholders reject the legitimacy of the governing legal regime itself, the remaining dispute is political, not conceptual, and the engineering move does not resolve it.

Revision costs and displacement

• Rhetorical flexibility — assessment: lost. “Fairness” can no longer serve as a blanket defense for any outcome; challenged decisions must point to the specific pre-declared rule applied. This is a feature for some audiences and a cost for others.
• Essentialist stakeholder alignment — assessment: alienated. Those who believe “fairness” has a singular objective essence (meaning either strict colorblindness or proportional outcomes) will read the engineered definition as a corruption rather than an improvement.
• Cognitive load on hiring committees — assessment: increased. Committees can no longer rely on a vague fairness intuition and must justify decisions against explicit multi-variable criteria.
• External statutory constraints — assessment: imposes a hard boundary. The re-engineering cannot occur in a legal vacuum. Post-Students for Fair Admissions v. Harvard (US, 2023) and analogous constraints elsewhere foreclose or constrain race-conscious affirmative action. If transparent value-weighting violates strict-scrutiny or disparate-impact thresholds, the engineering fails practically regardless of philosophical coherence. This caps the weight assignable to bias-correction or representation and motivates a two-track structure: the full construct as the aspirational specification, plus a jurisdiction-clause constraining which components are operationally activatable under applicable law.
• Conceptual unity — assessment: lost. “Fairness” stops being a single intuition-pumping word; the rhetorical force of “we just want to be fair” is lost.
• Stakeholder solidarity — assessment: lost. The word currently unites people who mean opposite things by it; after disambiguation the policy is visibly committed and some constituencies will read it as having picked a side. This is a feature if a coherent policy is wanted, but a cost if coalition maintenance matters.
• The appearance of neutrality — assessment: lost on contested political questions.
• Apparent simplicity — assessment: lost. The revised policy is longer, more technical, and harder to summarize. Stakeholders wanting a one-line answer (“we hire the best person”) will find it unsatisfying even when its substance is what they want.
• Honest net assessment: The case for revision rests on the diagnosis that the inherited ambiguity serves no one, producing decisions defensible on no one’s concept of fairness. However, if the inherited ambiguity is genuinely working in context (managing stakeholders, deferring hard choices, absorbing criticism by redefinition), revision is the wrong move and the cost is not worth paying.

Confidence per finding

• Function-failure diagnosis: High confidence. The four tensions are well-documented in organizational sociology and employment law. There is medium-high confidence that the four-function list is exhaustive, though a fifth function (dignity, opportunity) may be silently present.
• Proposed revision(s): Medium-high confidence in the structured, explicit, auditable construct (whether role-weighted or priority-ordered-component). There is lower or contingent confidence in the specific priority order or per-role weightings, as these are value commitments, not logical necessities, and organizations may reasonably rank differently. Confidence in specific threshold values (like the 4/5ths rule) is also contingent, as it represents a defensible default rather than a uniquely correct choice. The structure aligns directly with function-first methodology.
• Adoption feasibility: High confidence in the abstract, given the bounded population and direct institutional lever to rewrite and operationalize the policy. In practice, feasibility is jurisdiction-dependent; under strict anti-preference regimes, the bias-correction and representational-monitoring components may be constrained to monitoring and disclosure only, representing a coherent but smaller substantive commitment than the aspirational construct. The residual gap regarding external readers applying their own inherited concept is real and unfixable.

Additional considerations

• Operationalization of weightings / priority order: Who sets role-specific weightings, what evidence justifies them, and how they are audited remain domain-specific gaps requiring internal policy design and subject-matter expertise. The priority order specified (process integrity floor → bias-correction → representational monitoring → merit) is defensible but not the only defensible order; resolving it requires explicit organizational values clarification by leadership.
• Pool-scoping methodology: Even with the 4/5ths rule specified, the relevant labor-market availability pool (geographic scope, qualification-tier weighting, incumbency vs. external pipeline) is itself an empirical construct; the policy commits to publishing its methodology but must own the choice. This would resolve with a commissioned empirical labor-market study.
• Jurisdictional definition / component activation: The operative shape of the construct depends on the legal regime (US EEOC / post-SFFA vs. UK Equality Act 2010 vs. EU directives). This would resolve with specification of the target legal framework(s) and alignment with local employment-law counsel.

Target concept

The concept on the table is fairness, used as a load-bearing term in your hiring policy. The engineering question: what function should fairness perform such that it stops collapsing under the four jobs it is asked to do simultaneously — equal treatment, equal outcomes, bias-correction, and meritocracy — and is that function better served by revising the concept, renaming it, or restructuring the policy around it?

Before the engineering proper, four corrections to how the question was originally framed, because each one shapes what a sound answer can claim:

• “Re-engineer rather than just describe” is a false either/or; sequence matters. Conceptual engineering that skips the descriptive step engineers against a strawman. The current policy’s use of the word must be mapped before redesign. Describe first, then engineer.
• Some of the diagnosed “incompatibility” may be mode-confusion, not real conflict. When two people say “that wasn’t fair,” one may report current usage (“our process wasn’t identity-blind”) and the other propose a revision (“it should have corrected for background”). That is two people in different modes talking past each other, not a contradiction in the concept. A careful descriptive pass dissolves some apparent tensions before any engineering is needed. (Confidence: high that describe-first is correct; medium on how much of the specific tension is mere mode-confusion — unverifiable without the policy text.)
• A confident value-ranking would be a smuggled stipulation. The tempting move is to ratify “you’ve diagnosed the right problem” and then supply a guessed ordering (“for most organizations… the job gets done well, then transparency, then…”). That ranking is the single most important input to the engineering, and it cannot be legitimately supplied from outside. Inventing it dresses a preferred conclusion as the method’s output. The legitimate work here is structural; the value-weighting is yours.
• Jurisdiction is load-bearing. The legal content below assumes US jurisdiction. Under the UK/EU Equality Act 2010 regime the viability of some candidate revisions flips (this recurs in the ameliorative purpose, the candidates, and the costs). If the policy is governed elsewhere, the legal-interface conclusions change before the conceptual ones do.

Current usage — descriptive baseline

This section maps what fairness currently does in the policy. It is descriptive, not evaluative; the ameliorative move does not begin here.

• fairness is a synonym-bundle, not a criterion. In practice “a fair process” is invoked to mean blind/consistent (same questions, same rubric), non-discriminatory (legally compliant), unbiased (no animus or drift), and deserved (best person won) — often in the same paragraph. Its working function is rhetorical closure: it signals “this decision is defensible” without specifying along which axis.
• The four jobs map to standard names, each carrying a commitment. Equal treatment = procedural/anticlassification (identity-blind, rule-governed, same process for all); bias-correction = anti-subordination/corrective (identity-conscious offsetting of structural disadvantage); equal outcomes = distributive (hire composition tracks pool/population); meritocracy = desert (most job-relevant-qualified candidate wins). The inherited word fuses all four because ordinary ‘fairness’ genuinely means all of these in different contexts — this is not sloppiness on the org’s part.
• fairness carries a legal commitment the org did not choose. US employment law has already partly engineered the concept: “fair” sits on top of two distinct tests — disparate treatment (intent-based, maps to “equal treatment”) and disparate impact (effect-based, maps to “bias-correction”). The disparate-impact test originates in Griggs v. Duke Power Co., 401 U.S. 424 (1971) — facially neutral practices with disproportionate effect are unlawful absent business necessity — codified by the 1991 Civil Rights Act. Griggs-style impact doctrine is corrective fairness already encoded in law. The concept therefore has external owners (courts, candidates, regulators); it is not a blank concept awaiting design.
• fairness carries a moral/brand commitment. “We hire fairly” is recruiting language and an employer-brand promise; it does affective and reputational work a coined internal term does not.

This baseline bounds the engineering: the org is revising a concept with external uses it does not control, not authoring one from scratch. (Confidence: high on the legal/usage structure; the specifics of your actual policy text are unavailable — the policy language, if it exists in retrievable form, is the one input that would sharpen the baseline.)

Identified function failures

Sorting genuine trade-offs from false dichotomies is itself part of the engineering value — the original four-way framing overstates how much real conflict there is:

• Equal treatment ↔ bias-correction — genuine, manifesting at the decision point. You cannot be identity-blind and identity-conscious about the same decision simultaneously; they collide at the disparate-treatment/disparate-impact seam (the point where a facially neutral criterion produces adverse impact). This is the anticlassification/anti-subordination divide. Qualification: it is a triggered collision rather than a permanent contradiction, and it may be constitutive — possibly not fully resolvable, only allocatable. Cost of the failure: the policy gives no rule for which reading governs when the seam is hit.
• Meritocracy ↔ equal outcomes — genuine and deep. Different starting distributions plus accurate point-in-time merit measurement cannot both yield proportional outcomes; merit selection over a skewed qualified pool reproduces the skew. This is the irreducible one. (The merit=fair equation is itself contested: Brian Carey, “Justice in Hiring: Why the Most Qualified Should Not (Necessarily) Get the Job,” J. Applied Philosophy 41(4):731–744 (2024), argues justice often calls for a lottery among the sufficiently-qualified — cited to mark the contest, not to endorse the lottery argument.) Cost of the failure: the deepest values choice is laundered into a question about what the word means.
• Bias-correction ↔ meritocracy — partly false, two moves conflated. Correcting measurement bias (noisy proxies, biased rubrics) improves merit-tracking — they align. Correcting background disadvantage departs from point-in-time merit — they conflict. The common framing treats these as one move; they are two. Cost of the failure: people argue merit-vs-diversity when the live dispute is whether a given rubric validly measures merit.
• Equal treatment ↔ meritocracy — a false dichotomy only conditionally. Uniform, job-relevant criteria serve both at once only when the criteria are themselves validated as job-relevant and unbiased. Uniform application of a socially-coded criterion (a credential proxying for class, a Griggs-style “intelligence test”) is procedurally equal yet not merit-tracking. The dichotomy collapses into harmony when criteria are validated and re-opens into genuine tension when they are not. Cost of the failure: “we apply the same standard to everyone” gets treated as settling merit when it may not.
• Bias-correction ↔ equal outcomes — distinct, routinely conflated. Correcting process bias does not entail targeting outcome parity; pursuing outcome parity is not the same as removing process bias. Cost of the failure: much real-world argument bogs down here because one word covers both.

Core failure: fairness cannot tell you which constraint governs when two are genuinely triggered against each other — it names all four and adjudicates none. Secondary failure: it hides the conditional/false conflicts behind the deep ones, so people argue merit-vs-diversity when the live dispute is actually whether a given rubric is a valid merit measure. Because the word fuses genuine and false conflicts, “was it fair?” is unanswerable — four yardsticks are held up at once.

Cost in practice: every contested hire re-litigates the meaning of the word instead of the merits of the decision; the policy can be cited by both sides of any dispute; the genuine trade-off (a values choice leadership should own) gets laundered into a definitional question that looks technical. The measurable proxy — the fraction of contested hires that re-litigate the word vs. the decision — is also the metric a staged rollout can instrument. (Confidence: high on this diagnosis.)

Ameliorative purpose

The purpose is not “produce more diverse hires” or “protect merit” — those are conclusions, and smuggling either in is the failure mode. Stated as a function, the revised concept(s) should let a user of the policy tell the four jobs apart, so that:

• Priority under collision can be adjudicated — when two constraints are genuinely triggered against each other, the concept points to which governs, by reference to a stated ordering.
• The trade-off is made visible at the decision point — surfacing which axis a decision is justified on, rather than closing discussion with “it’s fair.”
• The burden of justification is allocated — specifying who must explain what when an outcome departs from a given constraint.
• A clean interface to the legal tests is preserved — internal reasoning maps onto the operative legal regime without a lossy translation layer. Which tests is jurisdiction-dependent: US = disparate-treatment/disparate-impact (Title VII); UK/EU = the Equality Act 2010’s direct/indirect-discrimination tests plus “positive action” provisions, which permit some comparable-merit corrective steps US disparate-treatment doctrine forbids.

The statement is deliberately neutral on which job wins — that neutrality is what keeps it a function and not a smuggled stipulation. Critically, the first function requires an input the concept cannot generate: your organization’s actual priority ordering among merit, correction, transparency, and cost. The engineering produces the machine; leadership must supply the setting. (One candidate revision below — stage-partitioning — partially lowers, though does not eliminate, the need for a total ordering.)

Note: this concept’s contestation may be constitutive (essentially contested in Gallie’s sense); the engineering move risks treating as resolvable what is in fact the object of legitimate ongoing dispute. fairness/justice are Gallie’s paradigm essentially-contested concepts: part of what the concept is is that reasonable people durably, legitimately disagree about it — the contestation is not a defect to be cleaned up. This bounds the whole exercise. Within a single org you can legitimately stipulate — that is the valid move, and Cappelen would hold the engineering project justified regardless of broader uptake — but engineering does not dissolve the contest; it relocates it to the priority ordering (or, under the stage-partition candidate, to the stage assignments), which will itself be contested (by candidates who ranked merit higher, by advocates who ranked correction higher), and to the boundary between the policy and everyone reading it with the public concept intact. The honest claim is therefore narrower than “stop trying to define fairness and re-engineer it to track your values”: you can engineer an internal operating concept for decision-making, but you cannot claim to have settled fairness, and you should expect the political disagreement to re-appear one level down. Offering the revision with this caveat is the difference between a proposal-with-honest-residual and an overreach. Anyone claiming to have found what fairness “really should mean” for hiring has overreached. The candidate revisions below are offered with this caveat.

Candidate revisions

These four are preserved as genuine alternatives, not collapsed to one recommendation, because which is right depends on inputs only you can supply.

Candidate A: Disaggregate and retire the word. Replace fairness with named sub-concepts/mechanisms, each owning one job: e.g. eligibility/transparency (procedural) → opportunity/access parity (corrective, operating on barriers to entering the process) → job-relevant/bounded merit selection (merit), with any distributive consideration handled as an explicit, separately-justified tie-break or audit rule. Rationale: fresh terms don’t carry the inherited fusion, so each sub-concept is independently checkable and independently arguable. Tradeoffs and costs: there is a hidden stipulation that must be consciously endorsed — assigning three jobs their own mechanisms while demoting the distributive job to an audit/tie-break footnote is itself a value choice; the function statement promised “each job assigned to a distinct mechanism,” and A quietly ranks distributive concerns as secondary. If you do not want distributive considerations subordinated, A needs a fourth, co-equal mechanism (a composition-target rule), not a footnote. It also loses the word’s rhetorical and legal-interface value; an internal vocabulary nobody outside the policy team speaks gets quietly re-translated back into “fair/unfair” by managers and candidates; it is wordier and more bureaucratic-sounding. On burden-of-justification (function 3): not discharged automatically — disaggregation gives axes to argue on, but a burden rule must be bolted onto each level (e.g. “a selection-stage departure from rubric rank is documented by the panel chair”).

Candidate B: Keep fairness, but index it explicitly. Two flavours: (i) redefine fairness in the policy as “the decision satisfies our stated constraint ordering O” — openly relational; “fair” stops meaning “neutrally correct” and starts meaning “consistent with published priorities.” (ii) Retain the word but force a subscript — fair-as-to-process / fair-as-to-access / fair-as-to-selection — and ban the bare unindexed use. Rationale: keeps the affective/legal/legitimacy anchor while removing the false claim of neutrality and forcing disambiguation. Tradeoffs and costs: advertises that fairness is a choice, which some stakeholders read as an admission the process is “just your values” — politically costly even where honest; and indexed terms decay fast in practice (people drop the subscript and the fusion creeps back), making it weaker against semantic erosion than A. On burden-of-justification (function 3): discharged most naturally of the candidates — a published ordering O generates its own audit rule (“any departure from O is documented by the hiring manager with reference to which higher-priority constraint forced it”).

Candidate C: Lexical-threshold structure. A merit floor (qualified-to-do-the-job, non-negotiable) plus corrective mechanisms operating only among candidates above the floor (the “comparable merit, then prioritize X” tie-break). Rationale: cleanly separates the genuine merit↔outcome trade-off (handled by the floor) from the conditional ones (handled above it). Tradeoffs and costs: the tie-break step is where legal exposure concentrates (see costs below), and “comparable merit” is itself contestable, so it relocates rather than removes the argument. On burden-of-justification (function 3): needs an explicit burden rule bolted on to be auditable — the policy must require the panel to record that a tie was declared, on what criteria comparability was judged, and which tie-break value was applied. Without that rule the structure silently encodes a demographic preference (the legal hazard in costs).

Candidate D: Stage-partition / sequencing (engineer the procedure, not the concept). Rather than ranking the constraints, assign each to a distinct pipeline stage so they rarely collide head-on: sourcing governed by opportunity parity (widen reach, strip non-job-critical credential filters); screening governed by procedural consistency + measurement-bias correction (same rubric, validated criteria, calibrated scorers); selection governed by bounded merit. A stronger variant leaves fairness itself as an unrevised contested ideal and engineers only the documented decision procedure that allocates the trade-offs stage by stage. Rationale: collisions are dissolved by sequence rather than adjudicated by rank; the merit↔outcome trade-off is attacked upstream (improving the input pool) rather than at the selection step (a tie-break) where legal exposure sits; it respects the essentially-contested status of the word; and it lowers the total-ordering input the rest of this analysis flags as decisive, since most pairwise conflicts never co-occur in one stage. Tradeoffs and costs: it relocates rather than eliminates the deep merit↔outcome trade-off — correction moved upstream is slower, less visible, harder to attribute, and can read as “doing nothing” to stakeholders wanting a visible selection-stage lever; residual selection-stage ties still need some rule; and the unrevised word (in the stronger variant) keeps generating disputes around the edges of the procedure — it sidesteps rather than answers the conceptual question. On burden-of-justification (function 3): discharged stage-wise — each stage owner documents compliance with that stage’s single governing constraint, more auditable than a global ordering but creating seams at the hand-offs.

Non-exclusivity: these are not mutually exclusive — B can wrap C or D; D can sit underneath B’s published ordering; A with a D-style procedure underneath is coherent. The design space is preserved rather than collapsed to one recommendation.

Jurisdiction flips the menu: Candidate C’s tie-break is hazardous under US law but available under UK/EU positive-action provisions; Candidate D is the most jurisdiction-robust because it puts corrective weight upstream of the selection step where most legal exposure concentrates.

Which premise selects which candidate — a map reporting entailments (“if you believe X, then Y follows”), not recommendations; you supply the premise and the candidate selects itself:

• If the final-selection step must stay identity-blind (correction confined upstream) → Candidate A, distributive job demoted to audit-only.
• If fairness is doing legitimacy/coalition work you can’t afford to lose → Candidate B (keep the rallying term, pay the semantic-erosion cost instead of the legitimacy cost).
• If legal exposure is the dominating constraint → Candidate D’s procedure variant (leave the contested word legally undisturbed; put the engineering into an auditable, stage-vettable procedure).
• If you want correction to operate at the decision point, not just upstream → none of A/B/C/D as drafted fits; you need an A-variant with a co-equal corrective mechanism at selection, which sits closest to the legal fault line.

A worked example of the output shape, using one arbitrary, non-endorsed value-set — included to make the method concrete, explicitly not a recommendation: Illustrative premises: “a decision is wrong if the process was opaque, or if applicants faced unequal access barriers; merit governs the final pick; demographic composition is a diagnostic signal, not a target.” → Selects Candidate A → Filled concept: retire “fairness”; ship eligibility transparency + access parity + job-relevant selection; distributive composition becomes a quarterly audit flag triggering a review of whether access parity worked, never a tie-break → Named trade-off accepted: merit beats correction at the decision point; correction lives entirely upstream. Changing the first clause drops out a different candidate and fill. That substitution is the actual work, and it is yours.

Implementation problem

Cappelen’s core caution: proposing a revised meaning and getting a community to use it are very different things, and one has limited control over what words mean. Treating proposal as adoption is the trap.

The site is unusually tractable. A policy document plus training plus audit metrics is a genuine institutional lever — the bounded, partly-controllable kind the conceptual-engineering implementation literature treats as favourable (Sterken on linguistic intervention; the broader debate over who is well-positioned and possesses enough collective long-range control to make a revision stick). You are not shifting public usage of ‘fairness’ — Cappelen’s hard case — but usage inside a bounded population you partly control. That said, the adoption surface has four distinct constituencies with different leverage:

• Hiring managers / interviewers must reason in the new categories under time pressure; the default gravity returns to “did the best person win / was it fair.” Worse: if “fair” is redefined but the interviewer’s head still runs the inherited concept, the result is a homonym, not a revision, and audit data becomes noise because decisions are filed under a word people read differently. This is the strongest argument for Candidate A — retiring the word forces fresh terms with no inherited reading to fight. Without worked examples and rubric changes, the new concept stays on paper.
• Legal/compliance keeps operating in the governing regime’s language (US: disparate-treatment/impact; UK/EU: Equality Act tests) regardless of internal terms; an internally-engineered meaning that, e.g., licenses demographic tie-breaking can collide with the legal concept. If the engineered concept doesn’t map cleanly, you’ve added a translation layer, not removed one.
• Candidates, press, courts cannot be made to adopt the revised fairness. They use the inherited concept. This is the decisive constraint: you can engineer the operating concept inside the org, but the word keeps its public meaning at the boundary.

The mechanism that could carry it is not a memo (you can’t legislate meaning) — it is changed rubrics and scorecards (use forces meaning), training with concrete contested cases, and a deliberate two-vocabulary design (keep fairness as the external-facing word while using disaggregated terms internally). Prefer staged adoption rather than an all-at-once flip: pilot the disaggregated concept on one role family, instrument the “fraction of contested hires re-litigating the word vs. the decision” metric (and whether interviewers can name which axis a borderline call turned on), and expand only if disputes become more tractable. This gives an empirical feedback loop and surfaces the translation-layer and re-translation failures on a small surface first.

Revision costs and displacement

• Rhetorical/brand/legitimacy anchor — assessment: “We hire fairly” does recruiting, trust, and coalition work that “we run equity-constrained merit selection” cannot — the latter reads bureaucratic and can be heard as admitting you’ve abandoned fairness. The overloaded concept’s vagueness is partly a feature: it lets a coalition with different values all sign the same policy. Retiring the word has a real morale and employer-brand cost.
• Legal interface — assessment: the inherited concept, for all its mush, plugs directly into anti-discrimination doctrine. A coined internal vocabulary still has to be mapped back to the governing legal tests in any dispute — you don’t escape the legal concept, you add a layer over it.
• Shared vocabulary / coordination / simplicity — assessment: everyone currently “knows” what fairness means (vaguely); one word is easier to communicate, train, and audit than four mechanisms. A precise replacement is precise only for those trained in it — you trade a shared vague term for a private precise one, reducing communication until adoption is broad.
• The candidates’ concept is not yours to control — assessment: even with a perfectly engineered internal term, applicants judge you against their inherited ‘fairness’. Your stipulation governs your process, not their evaluation of it.
• A specific, jurisdiction-dependent legal cost — assessment: a tie-break of the form “prefer the candidate from an underrepresented group” is an explicit demographic preference at the selection step. Under US law the operative hazard is long-standing Title VII disparate-treatment doctrine, under which an explicit demographic preference at selection is legally exposed independent of any 2023 development. Students for Fair Admissions v. Harvard (decided June 29, 2023) is frequently cited but should be cited precisely: it is a Title VI education-admissions ruling that, per employment-law commentary (Jackson Lewis; Crowell & Moring), did not directly interpret Title VII. Its relevance to hiring is climate, not holding — it drove an uptick in “reverse-discrimination” claims and EEOC/state-AG attention, raising litigation risk without itself changing employment law. The hazard is real and pre-dates SFFA; the SFFA-to-employment extension is interpretive, not settled. Under UK/EU law the verdict inverts: the Equality Act 2010’s “positive action” provisions permit preferring a candidate from an under-represented group when candidates are equally qualified (a genuine tie), so Candidate C’s tie-break can be lawful there. Candidate C concentrates this exposure; Candidate D minimizes it by moving correction upstream. An engineered concept that quietly encodes a demographic tie-break can convert a values choice into litigation exposure in one jurisdiction while being entirely lawful in another. (Confidence: high that this is a live and jurisdiction-sensitive risk; not legal counsel — verify against governing law before encoding.)

Is the loss worth the gain? Defensible yes for the internal operating concept (disaggregation genuinely makes disputes tractable) and no for the public-facing word (keep fairness at the boundary). This two-layer answer is itself a finding that the original “you won’t want to call it fairness anymore” framing misses.

Confidence per finding

• Function-failure diagnosis: High. The concept genuinely fails to adjudicate collisions and hides conditional conflicts behind deep ones; the genuine trade-offs are fewer than “four incompatible jobs” suggests; the equal-treatment/meritocracy pair is false only when the criteria are validated as unbiased. Basis: structural analysis of the four jobs and their pairwise relations, robust to the missing policy text.
• Proposed revision(s): Moderate / low-to-medium, and conditional. The disaggregation move is sound; which candidate (A/B/C/D) is right depends on the published priority ordering and jurisdiction — inputs that cannot be supplied or invented here. Basis: the candidates follow from the diagnosis, but candidate selection is premise-dependent.
• Adoption feasibility: Low-to-moderate / low that uptake is clean, medium that the lever exists. Internal adoption is achievable through rubrics + training + a single-role-family pilot; external adoption of a revised fairness is essentially unavailable, arguing for a deliberate two-vocabulary design and against fully retiring the word. The interviewer-head (homonym) problem and the legal anchor are the two things most likely to defeat the proposal in practice. Basis: the conceptual-engineering implementation literature plus the four-constituency adoption map.

The decisive missing input

The one thing the method requires and that cannot be supplied from outside: your organization’s purposes, ranked, with the conflict-cases resolved. Concretely:

1. Complete “a hiring decision was wrong when ______” with every ending the org would endorse, then rank them.
2. For the genuine constitutive clash (identity-blind treatment vs. identity-conscious correction), state which governs the final selection and which is confined to upstream access — with the caveat that treating the upstream-access vs. final-selection split as the pivotal allocation is itself a substantive, rejectable design commitment: someone who wants correction to operate at the decision point would reject this framing and reorganize the whole design around a different axis.
3. State whether fairness is doing political/legitimacy work that can’t be lost (pushes toward B/D) or is free to retire (enables A).

That single case — not a definition — is what the engineered concept has to encode. Candidate D lowers but does not remove this requirement. If your actual policy language and the governing jurisdiction are available, the analysis moves from structural to a redline of the specific text.

Four uncertainties remain open and each names what would resolve it:

• Whether SFFA’s education-admissions holding transfers to employment selection is interpretive (the direct doctrinal basis is Title VII disparate-treatment, not SFFA), and whether the US is the operative jurisdiction at all. Resolves with: the org’s governing jurisdiction + employment-counsel review.
• The correct legal interface in ameliorative-purpose function 4 (disparate-treatment/impact vs. Equality Act 2010 positive action) depends on jurisdiction.
• Whether some of the specific “incompatibility” you diagnosed is mode-confusion rather than genuine conflict cannot be verified without the actual policy text.
• The adoption-feasibility calibration (“low it’s clean; medium the lever exists”) cannot be independently judged without employment-law and org-change domain input.