Chapter 5: The Moral Manifold

RUNNING EXAMPLE — Priya’s Model

Priya maps TrialMatch onto the 3×3 moral manifold. For the BEACON-7 match, μ = 1 (Empirical–Deontic): HealthBridge has a contractual duty to match using validated criteria—an observable fact. μ = 4 (Normative–Deontic): the algorithm ought to give rural patients a fair chance, but ‘fair’ is undefined in the codebase. μ = 6 (Normative–Consequential): matching more rural patients should improve trial generalizability. μ = 9 (Meta–Consequential): what theory of value grounds the assumption that statistical power is good? She fills in all nine cells and realizes TrialMatch uses at most three dimensions and ignores the other six entirely. Six dimensions of moral reality, invisible to her algorithm.

5.1 The Question of Base Space

Every tensor lives on a manifold. The stress tensor in materials science lives on the manifold of spatial points within a body. The metric tensor in general relativity lives on the manifold of spacetime events. The electromagnetic field tensor lives on the four-dimensional manifold of spacetime, and the gauge connection lives on a fiber bundle over it. The question for geometric ethics is: what is the manifold on which moral tensors are defined?

This is not a technical detail. The choice of base space determines:

What counts as a “point” in moral reasoning

What counts as a “direction” — what the tangent vectors are

What transformations are admissible and which quantities are invariant

What it means for two moral situations to be “nearby” or “far apart”

Where the smooth structure breaks down — where discontinuities, boundaries, and singularities occur

This chapter develops the concept of the moral manifold — the space of morally relevant configurations over which ethical tensors are defined. We build on the mathematical toolkit of Chapter 4, but proceed with philosophical care, distinguishing geometric structure from metaphysical commitment.

5.2 What Are the Points?

The first question is fundamental: what are the “points” of moral space? Several candidates present themselves, each with distinctive strengths and limitations.

Candidate 1: Possible Actions

On this view, the moral manifold M is the space of possible actions available to an agent. Each point x ∈ M represents a complete specification of what the agent does — not just “help the neighbor” but a full description including manner, timing, motivation, and foreseeable consequences.

Strengths: This connects directly to the practical question of ethics (“what should I do?”). The gradient of the satisfaction function, ∇S, points toward the best available action.

Weaknesses: Actions are agent-relative. My action space differs from yours. This makes interpersonal comparison difficult and global structure unclear. Moreover, the space may not be a manifold in any natural sense — discrete choices (accept or decline) do not admit smooth interpolation.

Candidate 2: States of Affairs

On this view, points are possible states of the world — complete descriptions of how things are or could be. Actions are then paths or vectors, not points: an action takes the world from one state to another.

Strengths: This is agent-neutral. All agents evaluate the same space. It connects to consequentialist intuitions: what matters is the state of the world, not who brings it about.

Weaknesses: The space is enormous — potentially infinite-dimensional. And it privileges states over processes, which deontologists and care ethicists will resist, since the moral significance of how one arrives at a state is lost.

Candidate 3: Situations

A situation is richer than a state: it includes not just how things are, but the relationships between agents, their histories, their commitments, and the options available. A situation is something like: “Alice has promised Bob X. Bob is in need of Y. Alice can provide Y but only by breaking the promise to Carol. Dana is observing.”

Strengths: This captures the morally relevant features without requiring a full specification of the universe. It is closer to how moral reasoning actually works — what matters is the structured configuration of agents, stakes, and relationships.

Weaknesses: What counts as “morally relevant” is theory-dependent. Different ethical theories may carve situation-space differently. The category of “morally relevant features” requires principled delimitation.

Candidate 4: Agent-Situation Pairs

On this view, a point is a pair (a, s): an agent a in a situation s. This allows agent-relative evaluations while maintaining a common framework.

Strengths: It captures both partiality and impartiality as different operations on the space — holding a fixed (what should I do in this situation?) vs. varying a (what should anyone do?). It connects naturally to the agent-indexed tensors of Chapter 6 and to the fiber bundle structure where agent-perspectives form the fiber over a base situation.

Weaknesses: More complex structure; requires specifying how the agent index interacts with the situation index.

Our Choice: Structured Situations

For the purposes of this book, we take the moral manifold M to be a space of structured situations — specifications of:

The agents involved and their relationships

The options available to each agent

The morally relevant features of the context (needs, promises, histories, stakes)

The epistemic state (what is known, by whom)

This is deliberately ecumenical. A consequentialist can project onto states; a deontologist can focus on the structure of relationships and commitments; a virtue ethicist can attend to the character of the agents; a care ethicist can foreground the relational structure. The manifold M is common ground; different theories correspond to different tensors, metrics, and contractions on M.

Definition 5.1 (Moral Manifold, Informal). The moral manifold M is a paracompact Hausdorff topological space (see Definition A.1 for the formal construction) whose points represent structured moral situations — complete specifications of agents, relationships, options, and morally relevant context.

Agent-perspectives enter as additional structure over M — specifically, as a fiber bundle whose base is the space of situations and whose fiber at each point is the space of agent-perspectives available in that situation. This separation of base from fiber allows us to ask independently: “What is the situation?” (a question about M) and “Who is evaluating?” (a question about the fiber). We develop this in Section 5.4.

[Modeling choice.] The moral manifold M is a model, not a discovery. Its justification is pragmatist: the nine-dimensional structure converges across cultures and languages (the cross-lingual evidence of §17.7 shows stable dimensional weights in 11 languages spanning 3,000 years), it arises from structural necessity (any framework satisfying the three axioms of Chapter 9 produces an equivalent space), and it generates testable predictions that scalar alternatives cannot (the order-effect and semantic-gate results of §17.9–17.10). The dimensionality — nine, not seven or twelve — is a modeling choice whose adequacy is empirical. If future evidence reveals a moral phenomenon that the nine dimensions cannot represent, the manifold must be extended. If evidence shows that some dimensions are redundant, it must be reduced. The modeling-choice label on Axiom 5.1 applies equally here: the manifold is posited, not proved, and its warrant is that it works.

5.3 The Dimension of Moral Space

The Dimensionality Question

Having established what the points of M are, we must ask: what is the dimension of M? That is, how many independent directions of moral variation exist at each point?

This is not a question the mathematics can answer by itself. It is a substantive question about the structure of moral reasoning — one that admits of empirical investigation. But we can approach it systematically.

The 3 × 3 Derivation

A principled argument for the dimensionality of the moral manifold proceeds from two independently motivated classifications.

Three ethical scopes capture whose interests are at stake:

Individual — the focal agent’s own stakes, autonomy, and epistemic position

Relational — the structured obligations, claims, and care relationships between specific parties

Collective — the aggregate effects on groups, institutions, and the environment

Three epistemic modes capture what kind of moral question is being asked:

What Matters — the substantive values at stake (welfare, rights, justice)

Who Decides — the procedural and agentive questions (autonomy, legitimacy, virtue)

What We Know — the epistemic conditions (certainty, privacy, evidence)

Crossing these two classifications yields a 3 × 3 matrix of nine cells:

Each cell identifies an independent dimension of moral evaluation:

Consequences and Welfare (Individual × What Matters): Who benefits, who is harmed, at what scale?

Rights and Duties (Relational × What Matters): What claims, obligations, and permissions exist?

Justice and Fairness (Collective × What Matters): Is the distribution of benefits and burdens equitable?

Autonomy and Agency (Individual × Who Decides): Are individuals free to make their own choices?

Privacy and Data Governance (Relational × What We Know): Is personal information appropriately protected?

Societal and Environmental Impact (Collective × What We Know): What are the aggregate and ecological effects?

Virtue and Care (Relational × Who Decides): What does this reflect about the agent’s character and relationships?

Procedural Legitimacy (Collective × Who Decides): Was the process fair, transparent, and accountable?

Epistemic Status (Individual × What We Know): How confident are we? What don’t we know?

Arguments for Necessity and Sufficiency

Necessity. Each cell of the matrix is logically independent of the others. Consequences can be positive while rights are violated (a beneficent paternalism). Rights can be respected while justice is denied (each individual treated correctly, but the aggregate distribution is skewed). Autonomy can be preserved while epistemic conditions are poor (free choice based on bad information). No cell is reducible to a combination of others.

Sufficiency. The claim is that these nine dimensions subsume existing ethical frameworks. The EU High-Level Expert Group’s seven requirements for trustworthy AI map onto the nine dimensions without residue. The Markkula Center’s five ethical lenses map onto them. Beauchamp and Childress’s four principles of biomedical ethics map onto them. This subsumption is not proof of completeness, but it provides strong evidence that the nine dimensions capture the space of considerations that diverse ethical traditions have independently identified as morally relevant.

Falsifiability. The completeness claim is falsifiable: produce a genuine moral consideration that is (a) independent of all nine dimensions and (b) not a combination of them. If such a consideration exists, the framework must be extended. This is the appropriate epistemic stance toward a structural claim: hold it provisionally, test it against cases, and revise if necessary.

Formal Statement

Definition 5.2 (Dimension of the Moral Manifold). The moral manifold M has nine independent dimensions at generic (top-stratum) points; lower-dimensional strata may have fewer active dimensions. In local coordinates, a point p ∈ M can be parameterized by

p=(x¹,x²,…,x⁹)

where x^μ (μ = 1, …, 9) represents the state of the situation along the μ-th moral dimension. The tangent space T_pM at a generic point is a nine-dimensional real vector space.

We say “at least” because domain-specific refinements may introduce additional dimensions (the medical context adds benefit scores, waiting times, and clinical parameters that expand the local coordinate description). And we say “at generic points” because at stratum boundaries, the effective dimension may drop — some directions of variation cease to be available, as when an option becomes forbidden.

The Varying Metric

A crucial consequence of the nine-dimensional structure: the metric on M is not constant. Empirical data from the Dear Abby corpus (Chapter 17) reveals that the relative weights of dimensions shift by context:

In family contexts, social bonds (Dimension 7) dominate: the metric assigns greater “length” to variations in relational care.

In workplace contexts, procedural legitimacy (Dimension 8) and fairness (Dimension 3) dominate.

In neighbor disputes, fairness (Dimension 3) and rights (Dimension 2) dominate.

This context-dependence of the metric is precisely what differential geometry is designed to handle. The metric tensor g_μν(p) is a field on M — it varies from point to point. The fact that it varies is not a complication but the central phenomenon: moral space is curved, and the curvature encodes how the structure of moral evaluation changes across contexts.

5.4 The Differential Structure

Charts and Atlases on M

Recall from Chapter 4 that a smooth manifold requires coordinate charts that overlap smoothly. For the moral manifold, a chart assigns numerical coordinates to a neighborhood of situations.

Example. In a neighborhood of a family dispute about holiday hosting, local coordinates might be:

x¹ = degree of welfare impact on the host (Dimension 1)

x² = strength of the host’s duty to extended family (Dimension 2)

x³ = fairness of the hosting distribution over years (Dimension 3)

x⁴ = host’s autonomy regarding their own time (Dimension 4)

x⁷ = quality of the care relationship with family members (Dimension 7)

Not all nine dimensions need vary independently in every neighborhood. Some may be locally constant (privacy may be irrelevant to this particular dispute), reducing the effective local dimension. This is the sense in which the moral manifold may have varying dimension across its strata — though within a single stratum, the dimension is constant.

The Tangent Space of M

At each point p ∈ M, the tangent space T_pM has basis vectors {∂/∂x^μ}_μ=1,…,n where n is the local dimension. Each basis vector represents an independent direction of moral variation:

∂/∂x¹ = “increase welfare impact, holding all else fixed”

∂/∂x² = “strengthen the duty, holding all else fixed”

∂/∂x³ = “make the distribution more equitable, holding all else fixed”

A moral vector — an obligation, a gradient of satisfaction, a direction of moral improvement — is a linear combination of these basis directions:

O=O^μ(∂)/(∂xμ)

The components O^μ give the strength of the obligation in each moral dimension. This is the fundamental advantage over scalar ethics: the obligation has direction, not just magnitude.

The Agent Fiber Bundle

Agent-perspectives enter as a fiber over M. At each situation p ∈ M, the fiber F_p consists of the available agent-perspectives in that situation — the physician, the patient, the family member, the institutional review board.

Definition 5.3 (Agent Bundle). The agent bundle is a fiber bundle (E, π, M) where the fiber F_p = π^{-1}(p) at each point p ∈ M is the set of agent-perspectives available in situation p.

Remark. The fiber F_p varies with the point p: different moral situations may involve different sets of available agent-perspectives. Consequently, the agent bundle E is, in general, a fibered space rather than a fiber bundle in the strict sense (the fibers need not be diffeomorphic to a fixed model fiber F). The bundle structure is locally trivial within each stratum — situations in the same stratum have the same structure of available perspectives — but may change across stratum boundaries.

A point of E is a pair (p, a) — a situation and a perspective. Agent-relative moral tensors are sections of tensor bundles constructed from E. For instance, the agent-indexed evaluation tensor M^μ_a(p) assigns, for each dimension μ and each agent a, an evaluation score at situation p.

Perspective shifts (Type 2 transformations, in the language of Section 5.5) are transformations on the fiber that leave the base point fixed. The correlative structure of Hohfeldian analysis — that Morgan’s obligation implies Alex’s claim — is a constraint relating different points on the same fiber: if M^2_Morgan = O (obligation), then M^2_Alex = C (claim). This constraint is a structural invariant of the fiber, not a consequence of any particular evaluation.

5.5 Coordinates and Admissible Transformations

The Problem of Multiple Descriptions

In physics, coordinate transformations have a precise meaning: a change from one coordinate chart to another covering the same region of manifold. The transformation laws for tensors (Chapter 4, Section 4.4) specify exactly how components change.

In ethics, we speak loosely of “different perspectives,” “different framings,” “different descriptions.” But not all of these are coordinate transformations in the geometric sense. Conflating them invites the criticism that geometric ethics is vacuous — that by choosing the right “coordinate system,” one can make any answer come out.

This section introduces a discipline. We distinguish three types of transformation, with progressively weaker geometric status.

Type 1: Coordinate Redescriptions

A coordinate redescription is a change in how we parameterize the same underlying situation. The situation itself — the point p ∈ M — is unchanged; only the coordinates (the labels, the representation) differ.

Example. Describing a resource allocation in terms of “amount to Alice” (x_A) vs. “amount to Bob” (x_B = T - x_A), where the total T is fixed. These are different coordinates on the same one-dimensional manifold.

Example. Describing an action as “withholding treatment” vs. “allowing natural death.” If these are genuinely synonymous — if they pick out exactly the same action in all morally relevant respects — then they are coordinate redescriptions.

The tensorial requirement. Any moral quantity that is a genuine feature of the situation must transform appropriately under coordinate redescriptions. Scalars are invariant; vectors transform by the Jacobian; rank-2 tensors transform by two copies of the Jacobian. This is the mathematical content of the Bond Invariance Principle:

Axiom 5.1 (Bond Invariance Principle). [Definition/Modeling choice.] If two descriptions d and d’ of a situation are related by an admissible coordinate transformation, then any legitimate moral evaluation must assign the same value to both: E(d) = E(d’). Equivalently, legitimate moral evaluations are gauge-invariant quantities on the moral manifold.

What this rules out: moral evaluations that depend on mere labeling. If “allowing to die” sounds better than “withholding treatment” but they denote the same action, a proper moral evaluation must not distinguish them. This is the ethical analogue of general covariance in physics.

What this does not rule out: evaluations that differ because the descriptions are genuinely different — because they pick out different features, different contexts, or different aspects of the situation. The framework demands invariance under re-description, not under re-situation.

In the language of group actions, BIP states that the moral evaluation functional E is constant on orbits of the admissible re-description group: E(g · d) = E(d) for all g ∈ G_adm and all descriptions d of a situation p ∈ M. Here g · d denotes the action of the re-description group element g on the description d (the fiber over p in the description bundle of §5.4), and G_adm is the group of admissible re-descriptions — the subgroup of Diff(fiber) satisfying the five admissibility conditions of §5.5. Chapter 12 will identify G_adm = D₄ × U(1)_H as the maximal such group consistent with axioms A1–A5. The BIP is thus a precise orbit-invariance condition, and the tensor transformation laws of §4.4 are the infinitesimal manifestation of this group action on the moral evaluation algebra.

Pre-flight Checklist: Bond Invariance Principle

Before applying BIP, verify each item:

Admissibility verified. Both descriptions pass all five conditions (§5.5): same physical grounding, same agent structure, same stratum, reversible, no evaluative loading.

Manifold located. Situation has explicit coordinates on nine dimensions and an explicit stratum.

Tensor rank declared. The moral quantity has a declared type; BIP requires appropriate transformation.

Gauge group scope. Transformation is within G_ethics = D₄ × U(1)_H (Thm 12.3).

Grounding adequate (AI). Ψ satisfies six adequacy conditions (Def. 18.1).

Type 2: Perspective Shifts

A perspective shift changes the evaluating agent while holding the situation fixed. This is not a coordinate change on M — it is a transformation on the fiber of the agent bundle (Section 5.4).

Example. Evaluating a kidney allocation from the physician’s perspective vs. a patient’s family’s perspective. The situation (who the patients are, what their conditions are) is the same. What changes is who is evaluating.

Tensorial treatment. Perspective shifts act on the agent index of agent-relative tensors. The rank-2 tensor M^μ_a (evaluation of dimension μ by agent a) transforms in the agent index a when we change perspectives. The framework makes no general demand that all perspectives agree — that would be an implausibly strong objectivity requirement. Instead, it demands that certain structural features are perspective-invariant:

The correlative structure: if one perspective assigns obligation, the correlative perspective must assign claim.

The constraint set: what is forbidden is forbidden from all perspectives.

The stratum structure: which moral regime applies is not perspective-dependent.

These invariants characterize the objective structure of the moral situation, while the perspectival variation (which dimensions each agent prioritizes, which trade-offs each is willing to make) is the legitimate space of moral disagreement.

Type 3: Theory Shifts

A theory shift changes the mathematical structure we impose on M: the metric, the connection, the constraint set, the admissible transformations themselves.

Example. Switching from a utilitarian metric (all dimensions commensurable, with trade-off ratios specified by the metric’s off-diagonal components) to a lexicographic metric (Dimension 3 has absolute priority; other dimensions matter only when Dimension 3 is tied). This is not a coordinate change; it is a change in the geometry of moral space.

Example. Switching from a theory that treats agent identity as morally irrelevant (strict impartialism) to one that gives weight to special relationships (moderate partiality). This changes which fiber-bundle structures are admitted.

Tensorial treatment. Theory shifts are not symmetries — they change the structure. Different theories correspond to different choices of metric g, connection ∇, constraint set C, and admissibility conditions. The geometric framework does not adjudicate between theories; it represents each theory precisely, making commitments explicit and disagreements localizable.

Summary: The Transformation Hierarchy

The discipline: When we say “moral claims should be invariant under redescription,” we mean Type 1 transformations. Types 2 and 3 are not symmetries in this sense — they are legitimate sources of variation that the framework makes explicit rather than concealing.

Admissibility in Practice: A Checklist

The BIP’s power rests on the boundary of “admissible.” A skeptic may object: if you control the equivalence relation, invariance is guaranteed trivially. The objection is correct in principle—which is why the boundary must be operationally defined and independently testable. Given two descriptions d and d^' of what is claimed to be the “same” moral situation, the following checklist determines whether d→d^' is an admissible (Type 1) transformation. All conditions must be met.

Same physical grounding. Would a sensor suite measuring welfare indicators, bodily-autonomy indicators, resource distributions, and other morally relevant observables produce the same readings under both descriptions? If not, the descriptions pick out different situations, and the transformation is not admissible.

Same agent structure. Do d and d^' identify the same agents, the same relationships, and the same power asymmetries? If a description obscures an agent (“mistakes were made” vs. “the CEO fired the whistleblower”), the transformation changes the agent index and is Type 2, not Type 1.

Same stratum. Do d and d^' place the situation in the same Hohfeldian stratum? If “allowing to die” evokes Liberty while “withholding treatment” evokes Obligation, the descriptions encode different moral regimes. This is detectable by checking whether the semantic-gate structure differs under the two descriptions.

Reversibility. Can a competent moral reasoner, given d^', reconstruct d without additional information? If the transformation is lossy—if it obscures morally relevant features—it is a projection, not a coordinate change. Projections are not admissible.

No evaluative loading. Does the transformation change the evaluative valence? “Freedom fighter” vs. “terrorist” is not a coordinate change; it smuggles in an evaluation along the normative-mode dimensions. Admissible transformations change coordinate labels, not evaluative content.

If all five conditions are met, the transformation is admissible (Type 1) and the BIP demands invariance. If any condition fails, the transformation is Type 2 or Type 3, and variance between the two evaluations is legitimate.

Tempting but Invalid Transformations

The following examples illustrate transformations that appear to be mere redescriptions but fail the admissibility checklist. Each is a case where naive application of the BIP would be a mistake.

1. “Killing” vs. “allowing to die.” These are often treated as synonyms describing the same physical outcome (a death). But they differ on the agent-structure criterion: killing requires an active causal agent; allowing to die involves the absence of intervention. They also frequently differ on stratum: killing may trigger the absorbing stratum for direct harm (Chapter 8, Type III boundary), while allowing to die may remain in the Liberty stratum. Checklist verdict: Fails criteria 2 and 3. Classification: distinct situations. BIP does not apply.

2. “Taxation” vs. “theft.” Both involve involuntary transfer of resources, and a libertarian critic may claim they are equivalent. But they differ on the procedural-legitimacy dimension ( x⁸ )—taxation operates through institutional authority; theft does not—and on evaluative loading: “theft” encodes a negative judgment that “taxation” does not. Checklist verdict: Fails criteria 2 and 5. Classification: Type 3 (theory shift). BIP does not apply.

3. “Surveillance” vs. “monitoring for safety.” Both describe the same physical activity (cameras recording people). The physical grounding may be identical. But “surveillance” carries negative evaluative loading (privacy invasion), while “monitoring for safety” carries positive loading (protective care). The transformation smuggles evaluation into the description. Checklist verdict: Fails criterion 5. Classification: Type 2 (perspective shift with evaluative content). BIP does not apply; the variance between the two framings is a legitimate signal of moral disagreement.

4. “He said / she said” perspective swaps. Describing a dispute from either party’s perspective may seem like a mere coordinate change. But it shifts the evaluating agent (Type 2), not the coordinate system: the power asymmetries, epistemic positions, and relational roles are redistributed. Checklist verdict: Fails criterion 2. Classification: Type 2. Variance is expected and informative.

5. “Economic growth” vs. “ecological destruction.” Both may describe the same industrial activity. But they foreground different dimensions of the moral manifold (welfare x¹ vs. societal/environmental impact x⁶), use different physical grounding (GDP metrics vs. ecosystem metrics), and carry evaluative loading in opposite directions. Checklist verdict: Fails criteria 1, 2, and 5. Classification: distinct framings invoking different regions of the moral manifold. The BIP correctly predicts that these should not yield the same evaluation.

These examples demonstrate that the admissibility boundary is not a free parameter that can be tuned to guarantee invariance. It is an operationally testable criterion rooted in the five-condition checklist. Failures of admissibility are detectable—they show up as violations of the agent-structure, stratum, or evaluative-loading conditions—and they are informative: they reveal that what appeared to be a mere redescription is actually a substantive moral reframing.

An Edge Case: Genuine Ambiguity

The five invalid transformations above were designed to fail clearly. But the checklist’s greatest value emerges when the verdict is ambiguous—when reasonable people might disagree about whether a transformation is admissible. Consider:

“Euthanasia” vs. “mercy killing” vs. “assisted dying.” All three describe a physician administering a lethal dose to a terminally ill, consenting patient.

1. Same physical grounding. Yes—the same patient, the same medication, the same outcome. ✓

2. Same agent structure. Yes—physician, patient, and family occupy the same roles. ✓

3. Same stratum. Ambiguous. “Mercy killing” may trigger the absorbing stratum for direct killing (§8.5), while “assisted dying” may remain in the obligation stratum (medical duty of compassion). “Euthanasia” is etymologically neutral (“good death”) but is culturally coded as a boundary term. This is where the checklist escalates.

4. Reversibility. Marginal—a competent reasoner can reconstruct any of the three from any other, but connotative weight is lost in translation.

5. No evaluative loading. “Mercy killing” loads positively (“mercy”) and negatively (“killing”) simultaneously. “Assisted dying” loads procedurally (“assisted”). “Euthanasia” is closest to neutral.

Checklist verdict. Criteria 3 and 5 fail for “mercy killing” → Type 2 or 3. Criterion 3 is ambiguous for “euthanasia” vs. “assisted dying” → the transformation’s admissibility depends on a governance decision about stratum assignment (§9.1).

The lesson. When the checklist produces ambiguity rather than a clear verdict, it has done its job. It has identified the precise locus of genuine moral disagreement—the stratum boundary at which “medical compassion” meets “taking a life”—rather than papering over it with a false equivalence. The equivalence relation ≈ is not subjective; it is decidable up to a governance parameter (the stratum assignment), and the checklist tells you exactly which parameter requires a decision.

5.6 Stratification

Why Stratification Is Essential

The most important structural feature of the moral manifold is not its smooth geometry but its stratification: the partition of M into regions (strata) where different rules apply, joined along boundaries where the rules change discontinuously.

Smooth manifolds, by definition, are uniform in dimension: every point has a neighborhood homeomorphic to ℝⁿ for the same n. But moral space is not uniform. There are regions of smooth trade-offs (within a stratum), hard boundaries (between strata), and singular points (where strata intersect pathologically). The moral manifold is not a smooth manifold but a stratified space in the sense of Whitney (Chapter 4, Section 4.8).

Strata, Boundaries, and Transitions

Definition 5.4 (Moral Stratification). *The moral manifold M admits a Whitney stratification {S_α}_{α ∈ A} where: 1. Each stratum S_α is a smooth manifold (possibly of different dimension). 2. Within each stratum, the moral evaluation function S and the moral metric g are smooth. 3. At stratum boundaries, S or g may be discontinuous, or the effective dimension may change.*

Threshold effects. Within the “consent” stratum, the degree of informed consent varies smoothly — a little more information, a little more understanding. But at the boundary between “sufficient consent” and “insufficient consent,” the moral evaluation jumps discontinuously. Below the threshold, the action may be impermissible; above it, permissible. This is a stratum boundary.

Categorical distinctions. The difference between a 17-year-old and an 18-year-old, between killing and letting die, between lying and remaining silent — these are boundaries where legal and moral rules change. Within each side of the boundary, smooth variation is possible. Across the boundary, the applicable framework shifts.

The forbidden region. Certain options are not merely low-value but excluded from consideration — moral non-starters regardless of other considerations.

Definition 5.5 (Constraint Set). *The constraint set C ⊂ M is a closed subset such that every satisfaction function S compatible with the stratification: M → ℝ ∪ {-∞} satisfies S_C = -∞. The boundary ∂C is a stratum boundary.*

Notation remark. The symbol S is used in two distinct senses throughout this work: (i) S: M → ℝ ∪ {−∞} denotes the satisfaction function (a scalar field on M), and (ii) S_α denotes strata of the Whitney stratification (subsets of M). Context always disambiguates: S with a subscript α ∈ A (the index set) refers to a stratum, while S alone or S(p) refers to the satisfaction scalar. Where both appear in the same expression, the satisfaction function is written S(p) or as the contraction I_μ O^μ to avoid ambiguity.

The constraint set represents absolute prohibitions: actions involving non-consensual harm to innocents, violations of inviolable rights, and the like. The framework does not determine what belongs in C — that is the work of normative ethics and democratic governance. But it represents the structure of constraints precisely: as a region with hard boundaries, distinguished from regions of trade-offs.

Nullifiers as Absorbing States

Empirical evidence (Chapter 17) reveals a special class of stratum boundary: nullifiers — conditions that override all other considerations and collapse obligations to zero.

The Dear Abby corpus (Bond 2026) identifies abuse as the primary nullifier: across all domain contexts (family, workplace, friendship, romance), the presence of abuse universally nullifies obligations (n = 582). Other nullifiers include danger (n = 218) and impossibility (n = 144, encoding the principle ought implies can).

In stratified-space language, nullifiers define absorbing strata — lower-dimensional strata from which there is no transition back. Once a situation enters the “abuse” stratum, the normal obligation structure collapses: the metric becomes degenerate along the obligation dimensions, and the previously operative duties cease to apply.

Definition 5.6 (Absorbing Stratum). A stratum S_0 ⊂ M is absorbing if, for any path γ entering S_0, the moral evaluation along γ is determined by S_0’s rules thereafter, regardless of the prior stratum’s structure.

Semantic Gates as Transition Functions

The transitions between strata are not arbitrary. They are triggered by specific features of the situation — what we call semantic gates, borrowing the language of quantum normative dynamics.

Example. The phrase “you promised” triggers a transition from the Liberty stratum to the Obligation stratum. The phrase “only if convenient” triggers the reverse transition, from Obligation to Liberty. These transitions are discrete — step functions, not sigmoids — and empirically measurable with high reliability (94% for “you promised,” 89% for “only if convenient”).

In the geometry of M, a semantic gate is a codimension-1 submanifold (a “wall” between strata) equipped with a transition function specifying how the moral evaluation changes upon crossing. The discreteness of these transitions — verified empirically — supports the stratified-space model over a purely smooth manifold.

5.7 Singularities and Dilemmas

A singularity in M is a point where the normal geometric structure breaks down. In general relativity, singularities are points of infinite density or undefined curvature — places where the theory reaches its limits. In geometric ethics, they are genuine moral dilemmas.

Definition and Examples

Definition 5.7 (Moral Singularity). A point p ∈ M is a moral singularity if: 1. The metric tensor g_p is degenerate (det(g_p) = 0), so that some directions of moral variation have zero “weight” and comparisons break down; or 2. The satisfaction function S is not differentiable at p, so that the gradient ∇S — the “direction of moral improvement” — is undefined; or 3. Multiple constraint surfaces intersect at p, creating a cone of forbidden directions that leaves no permissible path.

The three singularity types are: (i) metric singularities, where the moral metric degenerates and dimensional comparisons fail; (ii) gradient singularities, where the satisfaction function loses differentiability and the direction of moral improvement is undefined; and (iii) constraint singularities, where multiple prohibition surfaces intersect and the feasible region collapses. The singular set Σ ⊂ M is their union: Σ = Σ_metric ∪ Σ_gradient ∪ Σ_constraint. Each type has distinct geometric character: metric singularities are intrinsic (detectable from g alone), gradient singularities depend on S, and constraint singularities depend on the constraint set C.

Example (Sophie’s Choice). A mother must choose which of her two children will be killed. There is no “right” answer. Any choice involves irreducible moral loss. The situation is singular: the satisfaction function S has no global maximum in the feasible region, or has multiple local maxima that are genuinely incomparable — the metric between them is degenerate.

Example (Competing absolute obligations). A physician has promised confidentiality to a patient who reveals an intent to harm a third party. The obligation of confidentiality and the obligation to prevent harm are both operative, both point in opposite directions, and neither can be subordinated to the other without remainder. The gradient ∇S is undefined: there is no single “direction of improvement.”

Dilemmas as Features, Not Bugs

Singularities are not failures of the framework. They are features. The framework represents genuine dilemmas as singular points, rather than forcing a false determinacy. This is an advance over scalar approaches, which must either deliver an answer (by arbitrary tie-breaking) or fall silent (by declaring the situation beyond their scope).

The geometric representation is more precise: we can say why the dilemma is a dilemma — which constraint surfaces intersect, which metric components become degenerate, which dimensions of the moral tensor conflict. We can characterize the type of singularity (metric degeneracy vs. gradient failure vs. constraint intersection) and the moral residue that any resolution leaves behind (Chapter 15).

5.8 Topology of the Moral Manifold

Local and Global Structure

The differential structure developed so far is local: it describes the neighborhood of a point. The topology of M concerns its global properties — properties that are invariant under continuous deformation but may be invisible from any single neighborhood.

Connectedness. Can any two moral situations be reached from each other by a continuous path through M? Or are there disconnected components — perhaps corresponding to genuinely incommensurable moral frameworks, or to domains of life so different that no smooth interpolation between them exists?

The evidence suggests that M is connected within a single cultural-historical context: any two everyday moral situations can be related by a sequence of small modifications. But whether M is connected across cultures and historical epochs is a substantive question. The temporal stability data from the Dear Abby corpus (Chapter 17) suggests that the core topology of M is stable over at least three decades, though the metric varies.

Compactness. Is M compact (closed and bounded) or does it extend indefinitely? The constraint set C introduces boundaries, but does M have a “ceiling” — a maximum possible stakes, a highest conceivable obligation?

In practice, moral evaluation operates on a bounded region: the stakes in any real situation are finite, the number of agents is finite, and the morally relevant features take values in bounded ranges. We can model the relevant portion of M as compact without great loss.

Fundamental group. Does M have “holes” — non-contractible loops? A nontrivial fundamental group π₁(M) would mean that certain closed paths in moral space cannot be continuously shrunk to a point. This would have implications for holonomy (Chapter 10): transporting a moral evaluation around such a loop could yield a nontrivial change, even though the loop returns to the same situation.

Whether M has nontrivial topology is among the most intriguing open questions of the framework. The path-dependence of obligation (Chapter 10) provides indirect evidence: if obligations change when transported around loops, the manifold may have nontrivial holonomy, which constrains the topology. But settling this definitively requires the empirical program outlined in Chapter 29.

5.9 Two Worked Examples

Example 1: The Manifold of a Medical Allocation

Situation. A physician must allocate a scarce treatment among three patients (A, B, C). Each patient has a medical benefit score β_i ∈ [0, 1], a waiting time w_i ∈ [0, ∞), an age a_i, and a number of dependents d_i.

The manifold. The space of possible allocations is the 2-simplex:

M=Δ²={(p_A,p_B,p_C):p_i≥0, ∑_i^​ p_i=1}

where p_i is the fraction of treatment allocated to patient i.

Stratification: - Interior S₂ (dim 2): All patients have positive probability. Smooth trade-offs possible. - Edges S₁ (dim 1): Two patients share the treatment; one is excluded. Three edges. - Vertices S₀ (dim 0): Deterministic allocations. Three vertices. These are the actual decisions; the interior represents deliberation space.

This is a Whitney-stratified space with three levels. The frontier condition holds: each vertex is in the closure of each adjacent edge, and each edge is in the closure of the interior.

Constraint region. Suppose Patient C’s inclusion would constitute discrimination (the decision is based on a protected characteristic). Then the region p_C > 0 is forbidden:

C={(p_A,p_B,p_C)∈Δ²:p_C>0}

The feasible region is the edge from vertex A to vertex B — a 1-dimensional manifold.

Metrics. Different metrics on Δ² correspond to different ethical theories: - Euclidean metric: g = dp_A ⊗ dp_A + dp_B ⊗ dp_B. All movements equally costly. - Weighted metric: g_{ij} = δ_{ij}/β_i. Movement toward sicker patients (higher β) is “easier” — lower moral cost. - Lexicographic metric: Not a metric in the standard sense, but a limiting case: the allocation must first maximize benefit to the worst-off, then consider secondary criteria. This corresponds to a degenerate metric where the worst-off dimension has infinite weight.

Satisfaction function. S: Δ² → ℝ assigns a moral score to each allocation. On the interior, S is smooth. On the constraint boundary, S = -∞. The gradient ∇S, computed via the metric, points toward the morally preferred allocation.

Example 2: The Promise Landscape

Situation. Morgan has promised to help Alex move on Saturday. Various complications arise.

The manifold. The relevant region of M is parameterized by (at least) three dimensions:

x² = strength of Morgan’s obligation (Dimension 2: Rights and Duties)

x⁴ = cost to Morgan’s autonomy (Dimension 4: Autonomy)

x⁷ = quality of the Morgan-Alex relationship (Dimension 7: Care)

Strata. The promise creates two primary strata:

Obligation stratum S_O: The region where Morgan has an obligation to help. The promise is binding, the cost is manageable, and no nullifiers apply.

Liberty stratum S_L: The region where Morgan is free to decline. Either the promise was conditional, or a nullifier applies, or the cost has become prohibitive.

Semantic gates. The boundary between S_O and S_L is crossed by specific triggers:

“Only if convenient” → O → L transition. The obligation collapses. The Dear Abby corpus measures this gate at 89% effectiveness.

“You explicitly promised” → L → O transition. The obligation crystallizes. Measured at 94% effectiveness.

Abuse by Alex → O → L (nullifier). Universal across contexts. The stratum boundary is absorbing.

Path-dependence. Consider two paths from an initial state (Morgan made the promise) to a final state (Morgan is aware that Alex was rude last week):

Path α: Morgan learns Alex was rude, then recalls the explicit promise.

Path β: Morgan recalls the explicit promise, then learns Alex was rude.

If the final obligation differs depending on which information came first, the moral manifold has curvature in this region. The non-commutativity of these information-updates — learning about rudeness and recalling the promise do not commute — is precisely the non-abelian structure formalized by the D₄ dihedral group (Chapter 12).

5.10 The Manifold as Common Ground

The moral manifold M is the base space over which all ethical tensors are defined. Its points are structured situations; its structure includes a nine-dimensional tangent space (at generic points), a context-varying metric, a Whitney stratification into moral regimes, and singular points representing genuine dilemmas.

Different ethical theories correspond to different structures on M:

Different metrics (utilitarian vs. egalitarian vs. lexicographic)

Different constraint sets (what is absolutely forbidden)

Different contractions (how multi-dimensional evaluation reduces to action)

Different connections (how obligations are transported across contexts)

But all theories share M as common ground. This makes disagreement tractable: we can ask whether two theories differ in their metrics, their constraints, or their contractions. We can identify where they agree (perhaps on the constraint set — both agree that torture is forbidden) and where they diverge (perhaps on the metric — they disagree about the trade-off rate between welfare and equality). Disagreement, in this framework, has address.

The moral manifold is not the whole of ethics. It is the stage on which ethics plays out — the space of possibilities that moral reasoning navigates. The next chapter develops the actors: the tensors of various ranks that live on M — obligations as vectors, interests as covectors, the moral metric, and the fundamental contraction S = I_μ O^μ that yields satisfaction.

But without the stage, there is nowhere for the actors to stand.

Technical Appendix: Formal Definitions

Definition A.1 (Moral Manifold, Formal). A moral manifold is a paracompact Hausdorff topological space M equipped with: 1. *A Whitney stratification {S_α}_{α ∈ A} into smooth manifolds of (possibly) varying dimensions; 2. A partial order ⪯ on A satisfying the frontier condition: S_α ∩ cl(S_β) ≠ ∅ implies α ⪯ β; 3. Whitney’s condition (B) at all stratum boundaries; 4. A smooth structure on each stratum S_α making it a smooth manifold of dimension d_α.*

Definition A.2 (Coordinate Chart on M). A coordinate chart on a stratum S_α is a diffeomorphism φ: U → V from an open set U ⊂ S_α to an open set V ⊂ ℝ^{d_α}. A coordinate transformation is a composition φ’ ∘ φ^{-1}: ℝ^{d_α} → ℝ^{d_α}, required to be a diffeomorphism.

Definition A.3 (Admissible Transformation). An admissible transformation of the moral manifold is a stratification-preserving homeomorphism (restricting to a diffeomorphism on each stratum interior) ψ: M → M satisfying ψ(S_α) ⊂ S_{σ(α)} for some bijective, order-preserving map σ of the index set A. Type 1 transformations are admissible; Type 3 transformations are not.

Definition A.4 (Agent Bundle). The agent bundle over M is a fiber bundle (E, π, M) where: 1. The fiber F_p = π^{-1}(p) at each p ∈ M is the set of agent-perspectives available in situation p; 2. The structure group G acts on fibers by perspective transformations; 3. Correlative constraints (O_a ↔ C_b for all Hohfeldian correlative pairs (a, b) at p) are structural invariants of the G-action.

Definition A.5 (Moral Singularity). A point p ∈ M is a moral singularity if: 1. The metric tensor g_p is degenerate: det(g_p) = 0; or 2. The satisfaction function S is not differentiable at p; or 3. Multiple constraint surfaces of codimension 1 intersect transversally at p, and the feasible cone at p has no interior.

Singularities represent genuine dilemmas: points where the moral structure does not determine a unique best response.

Definition A.6 (Absorbing Stratum). *A stratum S_0 is absorbing if there exists a continuous retraction r: N → S_0 from a tubular neighborhood N of S_0 such that, for any moral evaluation function S compatible with the stratification, S_N factors through r: S(p) = S(r(p)) for all p ∈ N.* Additionally, S₀ is a trapping region: for any continuous path γ with γ(t₀) ∈ S₀, we have γ(t) ∈ N ∪ S₀ for all t ≥ t₀. This trapping condition reconciles the formal definition with the informal path-based characterization (Definition 5.6): once a moral trajectory enters an absorbing stratum, the evaluation is governed by that stratum's rules thereafter.

❖

The manifold is the ground. The tensors are the figures.

Before we can say what obligation points toward, we must know the space in which it points.

This chapter has constructed that space: M, the moral manifold — nine-dimensional, stratified, equipped with a context-varying metric, and rich enough to serve as common ground for every ethical theory that has a claim to completeness.