Chapter 9: The Origin of the Moral Metric — Discovery, Construction, or Governance?

RUNNING EXAMPLE — Priya’s Model

Who decided that Mrs. Voss’s 4.5-hour drive counts the same as a co-pay difference in TrialMatch’s scoring? Priya traces the answer: no one decided. The metric was learned from historical data—data reflecting which patients historically completed trials. Those patients tended to live near research hospitals. The learned metric assigns low weight to access burden, creating a feedback loop: rural patients are not matched, so they do not complete trials, so the model learns rurality predicts non-completion, so rural patients are not matched. This is metric capture by historical bias. The realist metric (medical suitability) is defensible. The constructivist metric (community agreement on fairness) was never consulted. The governance metric (democratic determination) was replaced by a training loop.

9.1 The Metric Is Not Given

Chapter 6 established what the moral metric does: it defines distances between moral states, angles between obligation vectors, inner products that allow comparison, and the trade-off structure that determines how values may be exchanged. The metric g_μν is the most consequential object in the geometric framework — more consequential, in a sense, than the obligation vectors or interest covectors, because it determines which comparisons are even possible. Two analysts who agree on every obligation and every interest can disagree about what should be done, if they employ different metrics. The metric is where the deepest moral commitments live.

But Chapters 5 through 8 deferred a question that can no longer be postponed: where does the metric come from?

In physics, the metric of spacetime is discovered empirically — it is a feature of the universe we inhabit, measurable through the behavior of light, clocks, and gravitational fields. Einstein’s field equations relate the metric to the distribution of matter and energy. The metric is not a choice; it is a fact about nature.

In pure mathematics, the metric is stipulated — we choose Euclidean, hyperbolic, or Riemannian geometry and explore the consequences. The choice is free; the consequences are binding.

In ethics, the situation is more complex. The metric encodes substantive moral commitments: which values are commensurable, at what exchange rate, and which are incommensurable. These commitments are not simply read off from nature (the universe does not tell us the trade-off rate between welfare and justice), nor are they arbitrary stipulations (some metrics are morally inadmissible). The metric is simultaneously too important to be arbitrary and too contentious to be discovered.

This chapter examines three traditional accounts of the metric’s origin and argues for a fourth.

9.2 The Realist Account: Discovery

The Claim

The moral realist holds that there are moral facts independent of our beliefs about them. On this view, the moral metric is discovered, not invented. Just as physicists discovered that spacetime has Lorentzian rather than Euclidean signature, moral inquirers discover the true structure of moral space — the actual trade-off rates between values, the genuine incommensurabilities, the real distances between moral states.

Formal statement. There exists a unique (or at most a small family of) metric tensor(s) g_μν^* on the moral manifold M such that all correct moral evaluations are consistent with g^*. The task of moral inquiry is to approximate g^*.

Strengths

Phenomenology of inquiry. Moral reasoning often feels like discovery. When we realize that a seemingly acceptable trade-off was actually illegitimate — that efficiency gains do not justify racial discrimination — we experience this as learning a fact about the metric: the off-diagonal component g_{efficiency, discrimination} is zero (or the discrimination dimension is part of the constraint set, not the trade-off space). Realism explains this phenomenology directly.

Moral progress. If the metric is a fact, then moral progress consists in getting closer to the true metric. The abolition of slavery, the recognition of women’s suffrage, the criminalization of child labor — these can be understood as discoveries about g^*: dimensions that were once treated as commensurable with economic benefit (human freedom, political participation, childhood welfare) were discovered to be incommensurable with it or to require absolute priority.

Error theory. Realism explains moral error: a person or society that trades welfare for injustice is wrong about the metric, in the same way that a physicist who uses a Euclidean metric near a black hole is wrong about the geometry of spacetime.

Challenges

Epistemology. How do we access the moral metric? In physics, the metric manifests through measurable phenomena — light bending, clocks slowing. What are the moral analogues? Moral intuitions are notoriously variable across cultures and epochs. If the metric is a fact, we lack reliable instruments for measuring it.

Metaphysics. What kind of entity is a moral metric? It is an abstract mathematical structure defined over a space of structured situations. Where does it exist? How does it relate to the natural world? The moral realist must either accept a Platonic ontology (the metric exists in an abstract realm) or find a way to ground it in natural facts — and the grounding project has proven extraordinarily difficult.

Disagreement. If the metric is discoverable, why do reflective, informed inquirers disagree about it? Persistent, apparently irresolvable disagreement between utilitarians and deontologists — who disagree about the metric’s off-diagonal terms and the presence of lexicographic priority — is difficult to explain if the metric is a straightforward fact awaiting discovery.

Geometric Gloss

If realism is correct, the moral metric g_μν^* is a feature of reality, and different ethical theories are competing hypotheses about g^*. The utilitarian hypothesis is g^*=diag(w₁,…,w₉) with specific weights. The Rawlsian hypothesis is that g^* has lexicographic structure. The incommensurability thesis is that g^* has zero eigenvalues. Moral inquiry is metric estimation.

9.3 The Constructivist Account: Agreement

The Claim

The constructivist holds that moral facts are constituted by the outcomes of idealized rational procedures. On this view, the moral metric is constructed through deliberation — perhaps behind a veil of ignorance (Rawls), through ideal discourse (Habermas), or via reflective equilibrium (Daniels).

Formal statement. The moral metric g_μν is the fixed point of an idealized deliberative procedure Φ:

g^*=Φ(g₀,I,C)

where g₀ is the initial metric (reflecting current moral commitments), I is the set of idealizing conditions (full information, impartiality, rationality), and C is the set of structural constraints (consistency, completeness, invariance).

Strengths

Agency. Constructivism connects the metric to rational agency. The metric is not imposed from outside; it is the output of our own rational capacities operating under ideal conditions. This grounds the authority of the metric in something we can understand: reason itself.

Procedure. Constructivism provides a procedure for resolving metric disagreements. When two parties disagree about the trade-off rate between welfare and rights, the constructivist says: submit the question to the idealized deliberative procedure. The outcome is authoritative because the procedure is fair.

Metaphysical modesty. Constructivism avoids mysterious metaphysics. The metric exists as a feature of rational agreement, not as a Platonic entity.

Challenges

Which procedure? Different constructive procedures yield different metrics. The Rawlsian procedure (maximin behind the veil) yields a lexicographic metric with priority for the worst-off. The Habermasian procedure (ideal discourse) yields a metric determined by the force of the better argument — but “better” is itself metric-dependent. The Scanlonian procedure (what no one could reasonably reject) yields yet another metric. If the metric depends on the procedure, and the choice of procedure is itself a moral question, constructivism faces a regress.

Idealization. Real deliberators never achieve ideal conditions. How close is close enough? And how do we know when we have approximated the ideal, if the ideal itself is what we are trying to determine?

Determinacy. Even under ideal conditions, deliberative procedures may not yield a unique metric. There may be multiple fixed points — multiple metrics consistent with all the idealizing conditions. If so, the metric is underdetermined by construction, and constructivism collapses into pluralism.

Geometric Gloss

The metric is the fixed point of a deliberative process. Different starting points or different idealizing conditions may yield different fixed points — moral pluralism at the level of metric choice. The constructivist metric is not unique but convergent: as deliberation proceeds under better conditions, the class of admissible metrics narrows. Whether it narrows to a point is the key open question.

9.4 The Expressivist Account: Projection

The Claim

The expressivist holds that moral claims express attitudes rather than describe facts. On this view, the moral metric is a projection of our evaluative sensibilities onto the world — a structure we impose, not one we discover or construct.

Formal statement. The moral metric is agent-indexed: g_μν(a) is the metric of agent a, reflecting a’s pattern of evaluative responses. “Objectivity” is intersubjective convergence — the overlap between individual metrics — not correspondence to a metric-independent fact.

Strengths

Motivational force. Expressivism explains why moral judgments motivate action. If the metric reflects my evaluative sensibilities, then recognizing that situation p requires action (that ∇S≠0 in my metric) is simultaneously recognizing that I care about the relevant dimensions. The motivation is built into the metric.

Diversity. Expressivism naturally accommodates the diversity of moral outlook across cultures, epochs, and individuals. Different agents project different metrics — and this is expected, not anomalous.

Metaphysical economy. No Platonic realm, no idealized deliberation. Just agents with attitudes, projecting structure onto the moral landscape.

Challenges

Structure, not just attitudes. The moral metric is not merely an attitude — it is a geometric object with mathematical properties (symmetry, non-degeneracy conditions, transformation behavior). Can expressivism account for this structure? An attitude might be “I care more about welfare than justice.” The metric encodes not just this ranking but the precise trade-off rate, the coupling between dimensions, the degeneracies, and the transformation law under change of coordinates. It is unclear that attitudes have this kind of structure.

The Frege-Geach problem. Moral claims embed in complex logical contexts: “If torturing innocents is wrong, then getting others to torture innocents is wrong.” Expressivist semantics for embedded moral claims remains difficult. The metric framework sharpens the problem: how do we express “if g₁₂=0 (welfare and justice are incommensurable), then…” in expressivist terms?

Convergence. If the metric is merely projected, why is there as much moral agreement as there is? The convergence of diverse moral traditions on certain structural features — the obligation/liberty distinction (which transfers cross-linguistically at 100% in BIP experiments), the correlative symmetry of Hohfeldian states, the universality of nullifiers — is difficult to explain if the metric is pure projection. Something structural appears to constrain the projection.

Geometric Gloss

The expressivist metric is subjective: g_μν(a) varies with the agent. Intersubjective moral facts are features of the overlap between agent-metrics. The set of “objective” moral claims is the set of claims invariant under all agents’ metrics:

Objective claims=⋂_a{c:c is true under g(a)}

If this intersection is large, moral objectivism is approximately correct. If it is small, moral relativism wins. The size of the intersection is an empirical question.

9.5 The Governance Account: Democratic Determination

The Proposal

I propose a fourth account suited to the geometric framework: the moral metric is neither discovered, constructed by idealized reason, nor projected by individual sensibility. It is governed — the output of legitimate governance processes that have authority to determine, for a community, how moral trade-offs are to be structured.

Formal statement. The moral metric g_μν for a community C is determined by a governance process Γ_C:

g_μν=Γ_C(inputs)

where the inputs include empirical data about the community’s moral commitments, the structural constraints of §9.6, the outcomes of legitimate deliberative processes, and the community’s constitutional commitments.

What Governance Means

The governance account occupies a specific position in logical space, distinct from the three traditional accounts:

The governance account holds that:

The metric is a governance artifact — like a constitution, a legal code, or a set of regulatory standards. It is neither a natural fact nor a subjective projection. It is a socially authoritative determination of how trade-offs are to be structured.

Different communities may legitimately adopt different metrics. This is not relativism — some metrics are inadmissible (§9.6) — but it is pluralism: within the space of admissible metrics, legitimate governance processes may select differently, reflecting different histories, values, and circumstances.

The metric can be revised. If a community’s moral understanding changes — if formerly acceptable trade-offs come to be seen as illegitimate — the metric can be updated through legitimate institutional processes. This is what moral reform is: a change in the collectively authoritative metric.

Disagreement about the metric is political as well as philosophical. Whether welfare can be traded against justice at a particular rate is not merely a question for moral philosophy. It is a question for democratic deliberation, institutional design, and constitutional law.

Why Not Relativism?

The governance account might seem to collapse into relativism: if the metric is whatever a community says it is, then anything goes. But this objection misunderstands the account in two ways.

First, the metric is constrained (§9.6). Not every metric is admissible. Some structural constraints — non-degeneracy on protected dimensions, internal consistency, invariance under admissible transformations — are non-negotiable. These constraints narrow the space of admissible metrics, ruling out metrics that encode arbitrary discrimination, internal contradiction, or dependence on morally irrelevant features.

Second, governance processes can be better or worse, more or less legitimate. A metric adopted through transparent democratic deliberation, informed by the relevant moral considerations, responsive to the claims of affected parties, and consistent with constitutional commitments, has more authority than a metric imposed by fiat or adopted through corrupt process. The governance account does not say “any metric is as good as any other.” It says: “the authority of a metric derives from the legitimacy of the process that produced it.”

The Analogy to Legal Systems

The governance account draws on an analogy with law. Legal systems determine the structure of permissible actions: what counts as property, what counts as a contract, what constitutes a crime. These are not natural facts (different legal systems answer differently), nor are they mere conventions (they carry real authority and have real consequences). They are governance artifacts — socially authoritative determinations backed by legitimate institutional processes.

The moral metric is analogous. It determines the structure of permissible trade-offs: whether welfare can be exchanged for justice, at what rate, and under what conditions. These determinations are not natural facts, and they are not mere preferences. They are governance decisions with real moral authority.

The analogy extends further. Just as constitutions constrain ordinary legislation (some laws are unconstitutional, regardless of popular support), there are meta-constraints on the moral metric (§9.6) that no governance process can override. These constraints play the role of moral constitutionalism: they define the boundaries within which legitimate metric choice operates.

Connection to the Framework

The governance account integrates naturally with the geometric framework:

The manifold M (Chapter 5) and the obligation fields O^μ (Chapter 6) are prior to the metric. They represent the structure of the moral situation and the duties it generates. These are facts about the situation, not governance choices.

The metric g_μν is determined by governance. It represents the community’s authoritative determination of how values trade off. Different legitimate governance processes may yield different metrics.

The contraction S=I_μO^μ is the interface between the factual and the normative. The obligation vector is factual; the interest covector reflects the governed metric; the satisfaction scalar is the verdict.

Stratification (Chapter 8) — the boundaries, nullifiers, and constraint surfaces — is partially factual and partially governed. Some boundaries are structural invariants (the Hohfeldian correlative symmetry, the universality of abuse as a nullifier). Others are governance choices (where to set the age of majority, the threshold for informed consent).

9.6 Constraints on Admissible Metrics

Not every metric is admissible. Even on the governance account, structural constraints narrow the space of legitimate choices.

Symmetry

The moral metric must be symmetric: g_μν=g_νμ. The trade-off rate between welfare and justice must be the same regardless of which direction the trade is made: if one unit of welfare can compensate for k units of justice, then one unit of justice must compensate for 1/k units of welfare.

This is a mathematical constraint (symmetric bilinear forms are the standard definition of a metric), but it has moral content. A metric where trading welfare for justice costs more than trading justice for welfare would encode an asymmetric privilege for one value over the other — not in magnitude (which is permissible and encoded in the diagonal components), but in the structure of exchange. This is ruled out.

Non-Degeneracy on Protected Dimensions

The metric should not be fully degenerate: there must be some directions along which comparison is possible. A totally degenerate metric ( g_μν=0 for all μ,ν) would make all moral comparisons impossible, rendering the framework vacuous.

More substantively, certain dimensions should have guaranteed weight: the metric should not assign zero weight to dimensions that a legitimate governance process has identified as morally relevant. If a community has determined that rights matter (Dimension 2 has positive diagonal weight), a governance revision that sets g₂₂=0 — effectively declaring that rights don’t exist — requires a higher level of constitutional justification.

Invariance

The metric must respect the admissible transformation structure (Chapter 5, §5.5). Under Type 1 transformations (coordinate redescriptions), the metric must transform as a (0,2)-tensor:

g_μν=(∂xα)/(∂xμ)(∂xβ)/(∂xν)g_αβ

This ensures that the trade-off structure does not depend on how the situation is described. The metric is a geometric object, not a coordinate artifact. The Bond Invariance Principle (Chapter 5, Axiom 5.1) requires this.

Non-Discrimination

The metric must not encode trade-offs that depend on morally irrelevant features — specifically, protected characteristics. If two situations differ only in a morally irrelevant feature (the gender of the patient, the race of the applicant), the metric must assign the same distances and inner products.

Formal statement. Let ϕ:M→M be a transformation that changes only morally irrelevant features. Then:

g_μν(ϕ(p))=g_μν(p) for all p∈M

The metric is invariant under morally irrelevant transformations. This is a constraint on the metric, not on the manifold: the manifold may include morally irrelevant coordinates (they describe the situation), but the metric must not depend on them.

Consistency with Structural Invariants

The metric must be consistent with the structural invariants identified in Chapter 8: the D₄ group structure of Hohfeldian transitions, the universality of certain nullifiers, and the cross-linguistic transfer of deontic structure. These invariants are not governance choices — they are features of the moral manifold itself. The metric must be compatible with them.

Example. The correlative symmetry O ↔ C requires that if the metric assigns a positive weight to the obligation dimension for agent a, it assigns a corresponding positive weight to the claim dimension for the correlative agent b. A metric that violates correlative symmetry is inadmissible.

Summary of Constraints

The space of admissible metrics is the space satisfying all five constraints. It is large — allowing genuine pluralism — but not unlimited. The constraints rule out the most egregious metrics (discriminatory, inconsistent, description-dependent) while leaving room for legitimate governance to determine the metric’s specific structure.

9.7 Learning the Metric from Data

The Empirical Project

A practical question: can the metric be learned from data? If we have access to moral judgments — ordinal comparisons, trade-off behavior, evaluations of situations — can we infer the underlying metric that rationalizes them?

This question connects the governance account to empirical ethics. The governance account says the metric is determined by legitimate institutional processes. But those processes produce outputs — laws, regulations, institutional decisions, advisory column verdicts — from which the operative metric can, in principle, be reverse-engineered.

Formal Setup

The metric learning problem. Given a dataset of moral judgments {(p_i,q_i,≻)}, where p≻q means “situation p is morally preferable to situation q,” find a metric g_μν such that:

d_g(p_i,ideal)<d_g(q_i,ideal) whenever p_i≻q_i

where d_g is the geodesic distance in the metric g, and “ideal” is the point in M representing the best achievable moral state.

This is the moral analogue of metric learning in machine learning: given similarity judgments, find the metric that reproduces them.

Evidence from the Dear Abby Corpus

The Dear Abby corpus — 20,030 real moral dilemmas with expert ground-truth responses, spanning 32 years (1985–2017) — provides a natural dataset for metric learning. Each letter presents a moral situation; the columnist’s response reveals an implicit judgment about the relevant trade-offs.

Analysis reveals several features of the operative metric:

Context-dependent weighting. The relative weights of moral dimensions shift by context (Chapter 5, §5.3): in family contexts, care (Dimension 7) dominates; in workplace contexts, procedural legitimacy (Dimension 8) and fairness (Dimension 3) dominate; in neighbor disputes, rights (Dimension 2) dominate. This is evidence that the operative metric is not constant but varies across the manifold — precisely the structure of a Riemannian metric field.

Temporal stability of structure. Despite variation in specific advice over three decades, the structural features of the metric — the relative ordering of dimension weights, the pattern of couplings, the location of stratum boundaries — remain stable. This suggests that while the metric’s parameters may drift, its qualitative structure is robust. The moral manifold has a stable topology, even if its geometry evolves.

Correlative symmetry. The correlative rates O ↔ C (87%) and L ↔ N (82%) are consistent across subperiods and context types. This is evidence of a structural invariant of the metric — a feature preserved under the variation that metric learning would discover.

Evidence from BIP Experiments

The Bond Invariance Principle experiments across 109,294 passages in 11 languages provide a different kind of evidence. The key finding: the deontic axis — the obligation/permission distinction — transfers at 100% across all languages tested, even when specific moral content fails to transfer.

Interpretation for metric learning. The deontic axis is a structural feature of the metric that is independent of language and culture. It is not learned from data — it is a constraint that data universally satisfies. This supports the claim (§9.6) that some features of the metric are non-negotiable structural invariants, while others are legitimately variable across communities.

The pattern is suggestive:

The structural invariants constrain the space of admissible metrics. The governance-variable features are the parameters that legitimate institutional processes determine.

Challenges for Metric Learning

Underdetermination. Many metrics may fit the same data. A finite dataset of ordinal comparisons does not uniquely determine a metric — it determines only an equivalence class of metrics that are ordinally equivalent. To pin down the cardinal structure (the actual trade-off rates), richer data are needed: not just “A is better than B” but “A is twice as much better than B as C is better than D.”

Noise. Human moral judgments are noisy — affected by framing, mood, order effects, and cognitive biases. The metric learned from raw judgments is contaminated by this noise. Denoising requires either very large datasets (which the 109K-passage BIP corpus provides) or structural assumptions (which the admissibility constraints of §9.6 provide).

Whose judgments? Aggregating moral judgments across persons reintroduces the metric problem at a meta-level: what is the metric for combining different people’s judgments about the metric? The governance account resolves this recursively: the combination procedure is itself a governance choice, subject to legitimacy constraints.

9.8 Pluralism and the Meta-Metric

The Question

If different communities legitimately adopt different metrics, is there a meta-metric for comparing metrics? Can we say that one community’s metric is closer to correct than another’s?

Three Positions

Metric monism. There is exactly one correct moral metric g^*, and all governance processes are attempts to approximate it. This is the realist position, transferred to the governance context. Its difficulty: it requires a standard independent of governance, which the governance account by design avoids.

Metric relativism. There is no basis for comparing metrics across communities. If community A’s metric permits a trade-off that community B’s metric forbids, neither is wrong — they are simply different. Its difficulty: it cannot account for moral criticism across communities, which is a central feature of moral life. Calling another community’s practices “wrong” (as we do with slavery, genocide, and systematic oppression) requires a standard that transcends the community’s own metric.

Structured pluralism. Metrics can be partially compared along several dimensions — but not totally ordered. There is no single meta-metric, but there are meta-criteria:

Internal consistency. A metric that violates its own structural constraints (e.g., a metric that encodes a trade-off between dimensions it elsewhere declares incommensurable) is defective by its own standards.

Coverage. A metric that assigns zero weight to an entire dimension of the moral manifold — ignoring, say, the environmental dimension entirely — has a deficiency: it cannot even formulate moral claims along the neglected dimension.

Invariance preservation. A metric that violates the structural invariants of §9.6 (correlative symmetry, non-discrimination, transformation consistency) is inadmissible by structural criteria, not by any community’s particular moral commitments.

Reflective stability. A metric is reflectively stable if the community that endorses it would continue to endorse it after full information and ideal deliberation. A metric adopted under conditions of ignorance or coercion is reflectively unstable.

Empirical adequacy. A metric is empirically adequate if it rationalizes the community’s considered moral judgments (not all judgments — considered ones, after removing distortions from bias, ignorance, and coercion).

These meta-criteria partially order the space of metrics without totally ordering it. Two metrics may each satisfy all five criteria while differing in their off-diagonal components — representing genuinely different but equally legitimate moral perspectives.

Formal Statement

Definition 9.1 (Partial Order on Metrics). Define the relation g≼g^' (metric g^' dominates metric g) if and only if g^' satisfies all five meta-criteria at least as well as g, and strictly better on at least one. This relation is a partial order on the space of admissible metrics.

The partial order has the structure of a lattice in favorable cases: given two metrics, there may be a join (the least upper bound — the weakest metric that dominates both) and a meet (the greatest lower bound — the strongest metric dominated by both). Whether the space of admissible metrics forms a lattice is an open question with substantial mathematical interest.

The Irreducibility of Metric Disagreement

Structured pluralism implies that some metric disagreements are irreducible: they cannot be resolved by appeal to meta-criteria, because both metrics satisfy all meta-criteria equally well. At this level, disagreement is genuinely political — it reflects different legitimate choices about how to structure moral trade-offs.

This is not a defect of the framework. It is a structural theorem: the space of admissible metrics, constrained by the criteria of §9.6, is not a singleton. Genuine moral pluralism exists at the metric level — and any framework that pretends otherwise is either concealing its assumptions or restricting the space of admissible moral perspectives too narrowly.

The geometric framework’s contribution is to make metric disagreement precise: we can say exactly where two metrics differ (in which components g_μν), what moral consequences follow from the difference, and which meta-criteria each metric satisfies. This is progress — not the progress of resolving all disagreement, but the deeper progress of understanding its structure.

9.9 The Metric as Moral Infrastructure

An Analogy

Consider a city’s transportation infrastructure: roads, bridges, transit lines, sidewalks. This infrastructure determines how people can move — which destinations are accessible, at what cost, by what routes. The infrastructure is neither a natural fact (cities don’t grow roads spontaneously) nor a subjective preference (my desire for a road doesn’t create one). It is a governance artifact: the output of institutional processes (city planning, legislation, funding decisions) with real consequences for how people live.

The moral metric is moral infrastructure. It determines how moral agents can reason — which values are comparable, at what trade-off rates, through what deliberative routes. It is neither a natural fact nor a subjective preference. It is a governance artifact with real consequences for how moral life is structured.

And like physical infrastructure:

It can be well-designed or poorly designed. A city with no sidewalks privileges drivers over pedestrians; a moral metric with zero weight on environmental impact privileges present benefit over future harm.

It is designed for a community. The infrastructure of Tokyo differs from the infrastructure of rural Montana, and both may be well-designed for their contexts. Similarly, different communities may have different moral metrics, both well-suited to their circumstances.

It constrains without determining. Infrastructure constrains which trips are easy and which are hard, but it does not determine where anyone actually goes. The moral metric constrains which trade-offs are legitimate and which are not, but it does not determine which actions anyone actually takes. The metric is the field on which moral agency plays out — not a substitute for agency.

It requires maintenance and revision. Infrastructure that was adequate in 1950 may be inadequate in 2025. A moral metric that was adequate before the development of powerful AI may be inadequate after. The governance account allows — indeed requires — periodic revision of the metric through legitimate institutional processes.

Implications for AI Systems

The governance account has direct implications for AI alignment. If the moral metric is a governance artifact, then AI systems that make morally significant decisions must:

Be calibrated to the community’s metric. An AI system operating in a community with a utilitarian metric should not impose a lexicographic one, and vice versa.

Be transparent about the metric. The operative metric should be inspectable — which dimensions are weighted, what trade-offs are permitted, where the constraint surfaces lie.

Be updatable. When the community revises its metric through legitimate processes, the AI system should be updated accordingly.

Respect the structural constraints. The invariance requirements (§9.6) apply to AI systems as much as to human moral reasoning. The Bond Invariance Principle requires that the AI’s moral evaluations be invariant under admissible redescription — a testable engineering requirement.

This is the connection between the philosophical question of this chapter (where does the metric come from?) and the engineering question of Chapter 18 (how do we build ethical AI?). The governance account provides the source of the metric; the geometric framework provides the form; the engineering implementation provides the mechanism.

9.10 Summary

The moral metric g_μν — the structure of permissible trade-offs between moral dimensions — is the most consequential and most contentious element of the geometric framework.

Four accounts of its origin:

The governance account is the most natural fit for the geometric framework. It locates the metric’s authority in the same place that the framework’s implementation must locate it: in the institutional processes that determine, for a community, how moral trade-offs are to be structured.

The space of admissible metrics is constrained by structural requirements — symmetry, non-discrimination, invariance, compatibility with structural invariants — but not reduced to a singleton. Genuine moral pluralism exists within these constraints, and the framework makes this pluralism precise rather than concealing it.

The metric is moral infrastructure: designed, revisable, community-specific within structural constraints, and consequential for how moral life is lived. The geometric framework does not provide the metric. It provides the vocabulary for articulating, comparing, constraining, and implementing whatever metric a community adopts through legitimate governance.

Technical Appendix

Definition A.1 (Admissible Metric). A metric g_μν on the moral manifold M is admissible if it satisfies: 1. Symmetry: g_μν=g_νμ . 2. Non-total-degeneracy: there exists a subspace V⊂T_pM of dimension ≥1 on which g|_V is non-degenerate. 3. Tensor transformation: g transforms as a (0,2)-tensor under Type 1 transformations. 4. Non-discrimination: g_μν(ϕ(p))=g_μν(p) for all morally irrelevant transformations ϕ . 5. Structural compatibility: g is compatible with the D₄ group structure and known nullifier constraints.

Proposition 9.1 (Non-Uniqueness of Admissible Metrics). The space of admissible metrics on M is non-empty and has dimension >0 . In particular, there exist admissible metrics that differ in their off-diagonal components while satisfying all five admissibility conditions.

Proof gg^'ϵh_μνϵg^'=g+ϵh▫ . The space Adm(M) is non-empty by Theorem 9.1(i), and has dimension ≥ 10 by Theorem 9.1(iii). In particular, the flat metric δ_μν and any perturbation δ_μν + ε h_μν (for symmetric h respecting the D₄ and U(1)_H constraints) are both admissible for sufficiently small ε > 0. These metrics differ on off-diagonal components, establishing positive dimension. □

Note on logical structure. This proposition’s proof references Theorem 9.1 (stated below). The logical dependence is not circular: Theorem 9.1 provides the explicit dimension count for Adm(M), which is used here only as an existence claim (Adm(M) is non-empty with positive dimension). The existence of admissible metrics can also be established directly: the flat Euclidean metric on ℝ⁹ satisfies all axioms, and small perturbations of it remain admissible.

Proposition 9.2 (Structural Invariants Constrain the Metric). If the correlative symmetry rates P(O↔C)=p₁ and P(L↔N)=p₂ are fixed structural invariants, then the diagonal components of the metric on the rights/duties dimension (Dimension 2) must satisfy:

g₂₂(a)=f(p₁,p₂)⋅g₂₂(b) for correlative agents a,b

where f is a function of the correlative rates. Perfect correlative symmetry ( p₁=p₂=1 ) implies f=1 — identical metric weights for correlative partners.

Proof abp₁baf(p₁,p₂)▫ . Under the correlative swap s_c: O ↔ C, L ↔ N, the expected satisfaction functional must be invariant: ⟨Σ⟩_{s_c} = ⟨Σ⟩. The proportionality factor f must satisfy: (1) f(p₁, p₂) = g₂₂(a)/g₂₂(b), which by the symmetry of the correlative exchange yields f(p₁, p₂) = p₁/p₂; (2) f(1,1) = 1 (perfect symmetry implies equal metric weights); and (3) f(p₁, p₂) · f(p₂, p₁) = 1 (applying the swap twice returns to the original). All three are satisfied by f(p₁, p₂) = p₁/p₂, and this is the unique solution consistent with the correlative symmetry constraint. □

Characterization of the Admissible Metric Space

The preceding propositions established that the space of admissible metrics is non-trivial. We now prove structural results that characterize this space precisely.

Lemma 9.1 (D₄ Constraint on Metric Components). [Conditional on Def. A.1, condition 5] Let g_μν be an admissible metric on M. The D₄ structural compatibility requirement forces g_{2μ}(O) = g_{2μ}(C) = g_{2μ}(N) = g_{2μ}(L) for all μ. This eliminates 27 degrees of freedom.

Proof. D₄ acts on the Hohfeldian positions via the 4-cycle r = (O C N L) and reflection s_c = (O C)(L N). Structural compatibility (Def. A.1, condition 5) requires g_μν(σ(p)) = g_μν(p) for all σ ∈ D₄. Since D₄ acts transitively on {O, C, N, L}, the orbit of O under ⟨r⟩ is the entire set, so g_{2ν}(O) = g_{2ν}(C) = g_{2ν}(N) = g_{2ν}(L). This collapses four independent values to one for each of nine metric components g_{2μ}, eliminating 9 × (4 − 1) = 27 degrees of freedom. □

Lemma 9.2 (U(1)_H Constraint on the Harm Dimension). [Conditional on Def. A.1, condition 5] The U(1)_H symmetry generates a Killing vector field ξ = ∂/∂x⁷ along Dimension 7. The Killing equation ensures that the metric is independent of x⁷ (∂7g_{μν} = 0 for all μ, ν). Combined with the product-structure interpretation of Axiom A5 — that the harm dimension is geometrically orthogonal to all other dimensions — this forces g_{7μ} = 0 for all μ ≠ 7. This eliminates 8 off-diagonal degrees of freedom.

Proof. The U(1)_H action is a continuous one-parameter group of isometries along Dimension 7. Its generator is the Killing vector field ξ = ∂/∂x⁷. The Killing equation ∇_μ ξ_ν + ∇_ν ξ_μ = 0 forces g_{7μ} = 0 for μ ≠ 7 (if g_{7μ} ≠ 0 for some μ ≠ 7, geodesics in the harm direction would couple to non-harm dimensions, breaking the independence of harm magnitude). This eliminates 8 off-diagonal components: g_{71}, g_{72}, …, g_{76}, g_{78}, g_{79}. □

Theorem 9.1 (Admissible Metric Space Structure). [Conditional on Def. A.1] Let Adm(M) denote the space of admissible metrics on the 9-dimensional moral manifold M. Then:

(i) Adm(M) is non-empty: the flat metric δ_μν satisfies all five conditions.

(ii) Adm(M) is a convex cone: if g, g’ ∈ Adm(M) and α, β > 0, then αg + βg’ ∈ Adm(M).

(iii) dim Adm(M) ≥ 45 − 27 − 8 = 10.

Proof. (i) Non-emptiness. The flat metric δ_μν satisfies all five admissibility conditions: symmetry, non-degeneracy (det δ = 1), tensor transformation, non-discrimination (δ is constant), and structural compatibility (δ is invariant under O(9) ⊃ D₄ × U(1)_H). (ii) Convex cone. For g, g’ ∈ Adm(M) and α, β > 0, αg + βg’ preserves all five conditions (symmetry, non-degeneracy, tensor transformation, non-discrimination, and compatibility are all linear or open conditions). (iii) Dimension bound. A symmetric 9×9 matrix has 45 independent components. D₄ eliminates 27 (Lemma 9.1). U(1)_H eliminates 8 (Lemma 9.2). Remaining: 45 − 27 − 8 = 10. □

Theorem 9.2 (Structured Pluralism). [Conditional on Def. A.1 and Def. 9.1] The partial order ≼ on Adm(M) is not total. That is, there exist admissible metrics g, g’ ∈ Adm(M) such that neither g ≼ g’ nor g’ ≼ g.

Proof gg^'ϵh_μνϵg^'=g+ϵh▫ . Construct two admissible metrics: g = δ_μν + ε e¹³ and g’ = δ_μν + ε e⁴⁵, where e¹³ and e⁴⁵ are symmetric perturbations coupling Dimensions 1–3 (welfare–justice) and Dimensions 4–5 (autonomy–care) respectively, with ε > 0 sufficiently small. Both are admissible by the convex cone property (Theorem 9.1(ii)). Consider a community C₁ where welfare–justice coupling is empirically supported, and a community C₂ where autonomy–care coupling is supported. Then g achieves higher empirical adequacy for C₁ while g’ achieves higher adequacy for C₂. Under any ordering respecting empirical adequacy, g and g’ are incomparable. □

This is not a deficiency of the framework but a theorem: genuine moral pluralism — the existence of incomparable but individually coherent moral evaluations — is a mathematical consequence of the admissibility constraints, not a philosophical assumption.

❖

The metric is not a datum and not a guess. It is a decision.

A decision made by institutions, constrained by structure, revisable by process, and consequential for every moral evaluation that operates within it.

This chapter has argued that the question “Which metric is correct?” is not a philosophical puzzle with a unique answer, but a governance challenge with structural constraints.

The geometric framework does not resolve the challenge. It gives us the language to face it honestly — to know exactly what we are choosing when we choose a metric, and what follows from the choice we make.