Chapter 15: From Tensor to Decision — The Philosophy of Moral Contraction

RUNNING EXAMPLE — Priya’s Model

The Board meeting arrives. Sarah Chen must present TrialMatch’s fairness assessment. She has Priya’s full tensor analysis: nine-dimensional obligation vectors, uncertainty matrices, metric concerns. The Board wants one slide. Sarah contracts: she picks a weighted sum, weighting medical dimensions heavily and access dimensions lightly. The resulting scalar looks acceptable. By the Scalar Irrecoverability Theorem, the Board cannot recover from this number the fact that rural patients are systematically excluded. Priya watches from the back of the room. The moral residue—what remains after contraction—sits heavy. Sarah’s contraction was technically valid. It was also a choice about which dimensions to sacrifice. That choice was never discussed.

15.1 The Moment of Choice

Throughout this book, we have built a tensorial picture of moral reality: multi-dimensional, perspectival, geometrically structured. Obligations point in specific directions. Interests weight dimensions. The metric encodes trade-offs. The stratified manifold marks boundaries where the rules change. Curvature makes evaluation path-dependent. Conservation laws constrain what re-description can alter. Superposition allows framings to coexist and interfere. Collective agency generates obligations that no individual bears.

This apparatus captures the structure of moral evaluation with a precision that scalar frameworks cannot match. But at the moment of decision, we must act — and action is one-dimensional. We do this, or we do that. The rich, multi-dimensional structure of moral reality must be compressed into a single scalar verdict: this option is best; this action is required; this policy is adopted.

The mathematical operation that performs this compression is contraction: the summation over paired indices that reduces tensor rank. The fundamental formula S=I_μO^μ (Chapter 6) is already a contraction — the pairing of an interest covector with an obligation vector to produce a scalar. But it is not the only contraction, nor is it uniquely determined by the tensorial structure. Different contractions yield different verdicts, and the choice of contraction is itself a moral commitment.

This chapter examines contraction philosophically. When is contraction necessary? Who chooses how to contract? What information does contraction discard? And what obligations persist after contraction — the moral residue of what was contracted away?

15.2 The Mathematics of Contraction

Contraction as Index Summation

Recall from Chapter 4 that contraction reduces tensor rank by identifying an upper index with a lower index and summing:

For a (1,1) -tensor T_ν^μ , contraction over μ and ν yields a scalar:

Tr(T)=T_μ^μ=∑_μ=1⁹ T_μ^μ

More generally, contracting a type (p,q) -tensor with a type (r,s) -tensor over k pairs of indices yields a type (p+r-k,q+s-k) -tensor. Each contraction reduces the total rank by 2 — one upper and one lower index consumed — and discards the information in those dimensions.

The Contraction Chain

The full chain from tensorial moral reality to scalar action-decision typically involves multiple contractions:

Step 1: Multi-agent to single-agent. The multi-agent evaluation tensor M_ia (Chapter 7, §7.3) is contracted over the agent index a — a social aggregation procedure that reduces the multi-perspective tensor to a single-perspective evaluation.

Step 2: Multi-option to single-option. The resulting vector S_i (one score per option) is contracted to a scalar verdict — selecting the option with the highest score, or the option above a threshold, or the option satisfying a lexicographic condition.

Step 3: Multi-dimensional to scalar. Within each step, the fundamental contraction S=I_μO^μ reduces the nine-dimensional moral structure to a single number.

Each step discards information. The multi-agent tensor preserves the full landscape of agreement and disagreement; the single-agent vector discards it. The multi-option vector preserves how close the alternatives were; the scalar verdict discards it. The nine-dimensional structure preserves which dimensions contributed; the scalar contraction discards it.

Non-Commutativity of Contraction

When multiple contractions are performed, the order may matter. Contracting over agents first and then over options may yield a different result from contracting over options first and then over agents.

Example from Chapter 7. Consider the evaluation tensor M_ia with three patients (i = Alice, Bob, Carol) and five perspectives (a = physician, families, committee). Contracting first over agents (social aggregation) and then selecting the maximum yields one verdict. Selecting the best option from each agent’s perspective first, and then aggregating the verdicts, may yield a different verdict — particularly when preferences are non-aligned.

This non-commutativity is not a bug in the framework. It is a feature: it makes visible the fact that how we aggregate perspectives and how we select options are independent moral choices, and that their order of application matters. Arrow’s impossibility theorem (1951) is, in this light, a result about the non-commutativity of contractions on preference tensors.

15.3 The Necessity of Contraction

Why Contract at All?

If contraction discards information — and it does, inevitably — why not maintain the full tensorial richness?

Practical necessity. At the moment of action, we cannot act on a tensor. We can only do one thing. The effector systems of moral agency — whether muscles, speech, resource allocation, or institutional machinery — require a single command. The kidney goes to one patient. The vote is cast one way. The policy is adopted or rejected. The world changes one step at a time, and each step is scalar.

Cognitive necessity. Even in deliberation, we cannot hold the full tensor in view. A nine-dimensional obligation vector with five agent perspectives has 45 independent components. The full moral tensor for a realistic situation — with uncertainty, multiple options, temporal structure, and collective dimensions — may have thousands of components. Human working memory holds roughly four items. Contraction is how we manage the gap between moral complexity and cognitive capacity.

Communicative necessity. When we justify our actions to others, we typically offer contracted reasons: “this was the best option,” “I had no choice,” “the benefits outweighed the costs.” Full tensorial explanations — “the obligation vector had components (0.3, 0.7, 0.1, …) and the interest covector weighted them as (0.2, 0.4, 0.1, …)” — are available in principle (and Chapter 18 argues that AI systems should provide them), but they are not how moral communication naturally works.

The contraction bottleneck. Between moral reality (tensorial) and moral action (scalar), there is a bottleneck. Contraction is how we pass through it. The bottleneck is not a deficiency of our framework or our cognition. It is a structural feature of the relationship between evaluation and action: evaluation is multi-dimensional; action is one-dimensional.

The Cost of Premature Contraction

If contraction is necessary, its timing is not. Contracting too early — collapsing the tensor to a scalar before all relevant information has been considered — discards information that might have changed the verdict.

Example. A policy analyst reduces a multi-dimensional impact assessment to a single dollar figure (cost-benefit analysis) before the assessment is complete. The contraction (monetization of all dimensions) discards information about distributional effects, rights implications, and environmental impacts that have not yet been quantified. The resulting verdict is not wrong — it is premature. It reflects a contraction performed before the tensor was fully populated.

The geometric framework counsels deferred contraction: maintain the tensorial structure as long as feasible, contract only when action requires it, and acknowledge what the contraction discards. This is a structural argument for a familiar practical wisdom: don’t simplify too early; don’t decide before you must; keep options open until the cost of openness exceeds the cost of closure.

15.4 Who Contracts?

The Agent

Often the individual agent chooses how to contract — which dimensions to weight, which perspectives to prioritize, which trade-offs to accept. This is the sense in which moral choice is “up to us” even when the tensorial structure of the situation is objective. Two agents facing the same obligation field, with the same interests at stake, may contract differently — and both may be acting permissibly, if the choice of contraction falls within the range of legitimate options.

Example. Two physicians, facing the same allocation decision, may weight urgency and prognosis differently (different interest covectors I_μ). The resulting contractions S=I_μO^μ yield different verdicts. Neither is necessarily wrong; they have made different — but both legitimate — contraction choices.

The Community

Social norms may specify how to contract. A community’s moral conventions — which considerations have priority, how competing values are balanced, what counts as “good enough” — are, in tensorial terms, a socially determined contraction procedure.

Example. In many communities, a promise creates a lexicographic priority: the duty to keep the promise (Dimension 2) is contracted first, and other considerations are relevant only if the promise is fulfilled or legitimately overridden. This is a community-level contraction norm — a shared convention about the order and weighting of index summation.

The Institution

Laws, regulations, and institutional rules perform contraction. A speed limit contracts the complex tensor of traffic safety, efficiency, freedom, enforcement cost, and environmental impact into a single number. A hiring rubric contracts the multi-dimensional evaluation of candidates into a ranked list. A sentencing guideline contracts the complex moral landscape of criminal justice into a range of acceptable punishments.

Institutional contraction has a distinctive feature: it is explicit. The contraction procedure is written down, debatable, and (in principle) revisable. This makes institutional contraction more transparent than individual or community contraction — and connects it to the governance account of Chapter 9. The governance account says: the metric is the output of legitimate institutional processes. The same applies to contraction: the legitimate contraction procedure is the one that emerges from legitimate governance.

The Algorithm

In AI systems, the contraction is specified by the designer, learned from data, or negotiated by governance. The choice of loss function is a choice of contraction: a loss function takes a multi-dimensional output space and produces a scalar to be minimized. Different loss functions embody different moral contractions.

Loss function=contraction of the moral tensor to a scalar optimization target

This observation is central to Chapter 18’s treatment of AI alignment. A misaligned AI is, in many cases, an AI performing the wrong contraction — optimizing a scalar that is a poor compression of the tensorial moral reality. Alignment, in this vocabulary, is getting the contraction right.

15.5 Types of Contraction

A Taxonomy

Different contraction methods embody different moral philosophies. The geometric framework makes these methods explicit and comparable.

Summative contraction. Σ the components with equal weight:

S=∑_μ^​ O^μ=O¹+O²+⋯+O⁹

This is classical utilitarianism at the individual level: all moral dimensions contribute equally. At the social level (contracting over agents), it is additive aggregation: total welfare is the sum of individual welfares.

Weighted contraction. Σ with unequal weights:

S=∑_μ^​ w_μO^μ,  w_μ≥0, ∑w_μ=1

This is pluralistic utilitarianism: different values have different weights. The weights w_μ are the components of the interest covector I_μ, so the fundamental formula S=I_μO^μ is a weighted contraction. Every ethical theory that yields a scalar verdict is performing a weighted contraction, differing only in the weights.

Maximin contraction. Take the minimum across a relevant index:

S=min_μO^μ or S=min_aM_ia

This is Rawlsian: the evaluation of a social arrangement is the evaluation from the perspective of the worst-off. The verdict depends only on the lowest component — the dimension (or agent) where the obligation is least fulfilled.

Lexicographic contraction. Order the indices by priority; contract by the highest-priority index first, and lower priorities break ties:

S=(O^π(1),O^π(2),…,O^π(9))_lex

where π is a priority ordering. This is deontological: some constraints have lexical priority. The obligation along the highest-priority dimension determines the verdict; lower-priority dimensions matter only in case of a tie.

Satisficing contraction. Contract to a binary — acceptable or unacceptable — by checking whether all components exceed a threshold:

S={[1, if O^μ≥τ_μ for all μ; 0, otherwise]

This is threshold deontology: options above the threshold are permissible; those below are not. The threshold vector τ=(τ₁,…,τ₉) is the moral minimum — the acceptable floor for each dimension.

Probabilistic contraction. Contract by expected value over uncertain indices:

S=E[I_μO^μ]=∑_μ^​ E[I_μ] E[O^μ]+∑_μ^​ Cov(I_μ,O^μ)

This is decision theory under uncertainty. The covariance term shows that when uncertainty in interests correlates with uncertainty in obligations, the expected satisfaction differs from the product of expected values. The uncertainty tensor Σ^μν of Chapter 6 (§6.6) enters here: the expected satisfaction depends on the alignment between the uncertainty structure and the interest-obligation pairing.

What Unifies Them

Each of these is a different mathematical operation on the same underlying tensor. The tensor does not change; the contraction does. The geometric framework reveals that much of moral disagreement reduces to disagreement about contraction rather than disagreement about the tensor itself.

This is a productive clarification. Two disputants who seem to disagree about “what should be done” may in fact agree about the obligation field O^μ, the interest covectors I_μ, and the metric g_μν, and disagree only about how to contract — whether to use summative, maximin, lexicographic, or some other procedure. This disagreement, once identified, is more tractable than an amorphous “moral disagreement.”

15.6 The Information Lost in Contraction

What Contraction Discards

A rank-2 tensor with n² components contracts to a scalar with 1 component. A nine-dimensional obligation vector contracts to a single satisfaction score. The ratio of information retained to information discarded is small. What, specifically, is lost?

Directional information. The tensor tells us which dimensions are engaged — which values are operative, which obligations are strongest, where the moral weight lies. The scalar tells us only the magnitude: “how good” or “how bad,” not “good along which axis” or “bad in what respect.”

Relational information. The tensor encodes relationships between dimensions — the off-diagonal elements of the metric, the correlations in the uncertainty tensor, the couplings between agents’ evaluations. The scalar discards all relational structure. Two situations with the same scalar score but very different tensorial structures are indistinguishable after contraction.

Perspectival information. The multi-agent tensor M_ia preserves the full landscape of agreement and disagreement. The scalar verdict discards it: we know the outcome but not who agreed, who dissented, or how close the alternatives were.

Near-miss information. The scalar verdict “give the kidney to Bob” says nothing about how close Alice was to being selected, or how much better Bob was, or whether Carol was a distant third or a competitive alternative. This information — the distance between the selected option and its alternatives — is morally relevant (a close call merits more caution and humility than a clear case) but is discarded by contraction.

An Information-Theoretic Measure

The information loss of a contraction can be quantified:

Definition 15.1 (Contraction Loss). Let T be a moral tensor of rank r and S=C(T) be its contraction to a scalar via contraction procedure C. The contraction loss is:

Remark. The contraction loss quantifies information destroyed by reducing a tensor to lower rank. It is defined in terms of the tensor’s rank and the contraction procedure, not as a Shannon or von Neumann entropy (which would require a probability distribution or density operator not present in the geometric construction). The appropriate analogy is the rank deficiency of a matrix projection: contracting a rank-k tensor to rank j < k discards the (k−j)-dimensional kernel of the contraction map. The loss functional L(π) measures the dimension of this kernel weighted by the magnitude of the discarded components.

L(C)=H(T)-H(S|C)

where H(T) is the entropy (information content) of the tensor and H(S|C) is the entropy of the scalar conditioned on the contraction procedure. The loss measures the information destroyed by the contraction.

Different contraction procedures have different loss profiles. Summative contraction loses directional information but preserves magnitude. Lexicographic contraction preserves the highest-priority information but loses everything else. Satisficing contraction loses all quantitative information but preserves the essential binary (acceptable/not).

The choice of contraction is, in part, a choice about which information to sacrifice. This is a moral choice, not merely a technical one.

The Scalar Irrecoverability Theorem

The information loss documented above is not merely quantitative — it is structurally irrecoverable. This result, which we call the Scalar Irrecoverability Theorem, provides the formal foundation for every domain-specific critique of scalar reduction (QALY-based health economics, utility-maximizing economics, scalar sentencing guidelines).

Theorem 15.1 (Scalar Irrecoverability). [Proved.] Let Q: ℝ⁹ → ℝ be any function that maps a 9-dimensional moral attribute vector to a scalar evaluation. Then: (i) Q is not injective: multiple morally distinct states map to the same scalar value. (ii) The information destroyed by Q is irrecoverable: there exists no function ψ: ℝ → ℝ⁹ such that ψ ∘ Q = id. (iii) The moral geodesic on the full manifold ℳ is in general different from the scalar-optimal path, and the divergence between them is not bounded by any function of Q alone.

Proof. (i) The rank-nullity theorem requires dim(ker dQ) ≥ 8 at every regular point of Q: a map from ℝ⁹ to ℝ has a differential of rank at most 1, so its kernel has dimension at least 8. Any two states differing only in a kernel direction map to the same scalar. Since the kernel is 8-dimensional at every regular point, the space of morally distinct states that are scalar-equivalent is generically 8-dimensional — not a set of measure zero but an 8-dimensional submanifold for each scalar value. (ii) By the data processing inequality, information that passes through a dimensionality-reducing map Q cannot be recovered by any downstream function ψ. Formally: H(α | Q(α)) > 0 whenever the conditional entropy of the attribute vector given its scalar image is positive, which it is whenever the kernel of dQ contains morally relevant variation — i.e., generically. (iii) Let γ*_ℳ be the geodesic on the full 9-dimensional manifold and γ*_Q be the path that maximizes Q. Since Q collapses 8 dimensions, it can assign equal value to paths with radically different moral profiles. The divergence between γ*_ℳ and γ*_Q depends on the structure of the moral manifold in the kernel of dQ, which is invisible to Q and therefore cannot be bounded by any function of Q alone. □

Remark (Domain Applications of Scalar Irrecoverability). The theorem has immediate consequences for any field that reduces multi-dimensional moral evaluation to a scalar. In health economics, the quality-adjusted life year (QALY) is a map Q: ℝ⁹ → ℝ that collapses clinical outcomes, autonomy, trust, dignity, justice, and epistemic status to a single number; Theorem 15.1 proves that the information destroyed is irrecoverable and that QALY-optimal policies can be manifold-suboptimal. In economics, scalar utility maximization discards the moral, social, and identity dimensions on which agents actually decide; the theorem explains why ‘irrational’ behavior (ultimatum game rejections, loss aversion, reference dependence) is rational on the full manifold. In law, scalar sentencing (years of imprisonment as the sole metric) discards rehabilitation, victim impact, community trust, and proportionality dimensions. Each of these is an instance of Theorem 15.1, not a separate disciplinary problem.

15.7 Moral Residue

What Contraction Leaves Behind

Even after the contraction is performed and the decision is made, the contracted-away information does not vanish from moral reality. It persists as moral residue — the normative significance of considerations that were present in the full tensor but did not survive into the scalar verdict.

Definition 15.2 (Moral Residue). The moral residue of a contraction C applied to a tensor T is the complement:

R=T-C^-1(C(T))

where C^-1(C(T)) is the tensor that, under contraction C, would yield the same scalar as T, with minimal norm. The residue is the part of T that contraction discards — the “shadow” of the contracted-away dimensions.

Moral residue is not merely theoretical. It manifests in concrete moral phenomena:

Regret. The feeling that something of value was lost in the decision, even if the right choice was made. A surgeon who amputates a limb to save a life has chosen correctly, but the loss of the limb is real. The residue is the component of the obligation tensor that pointed toward “save the limb” — a component that was contracted away in favor of the component pointing toward “save the life.”

Compensation. Obligations to make up for what was sacrificed. The patient not chosen for the scarce kidney (Chapter 7) retains a claim — perhaps to priority in future allocations, certainly to respectful explanation, possibly to alternative treatment. These claims are the moral residue of the contraction that selected another patient.

Acknowledgment. The duty to recognize what was set aside. A commander who orders soldiers into danger has contracted the tension between safety and mission success in favor of the mission. The residue — the safety consideration that was sacrificed — generates an obligation to acknowledge the sacrifice, to honor the fallen, to avoid casual disregard for what was lost.

Memory. The claim of the contracted-away on future decisions. A society that has resolved a moral dilemma by choosing one value over another (liberty over security, say) accumulates residue from the sacrificed value. This residue may build over time, eventually generating pressure to rebalance — a moral analogue of the buildup of stress in a geological fault, eventually released in a sudden shift.

Residue and Holonomy

Moral residue connects to the holonomy of Chapter 10. When an obligation is parallel-transported around a moral circuit and returns to its starting point, the holonomy — the rotation induced by the circuit — is a measure of the circuit’s residue. The moral experience has left a permanent mark: the obligation vector has been rotated, and the rotation encodes the residue of the experiences traversed.

The holonomy H_γ(O)-O is, in the language of this chapter, the accumulated moral residue of the trajectory γ. Each contraction along the trajectory discards information; the holonomy measures the cumulative effect of these contractions on the obligation’s orientation.

Residue Is Not Optional

Moral residue is not a psychological accessory to the decision. It is a structural feature of the contraction — an inevitable consequence of reducing a high-rank tensor to a low-rank verdict. Any contraction that discards information generates residue; the only question is whether the residue is acknowledged or ignored.

Proposition 15.1 (Inevitability of Residue). For any non-trivial moral tensor T (rank ≥1 ) and any contraction to a scalar, the moral residue R is nonzero unless T is already scalar — that is, unless there was nothing to contract.

Proof. If T has rank ≥1, it has at least 9 independent components (for a vector) or 81 (for a rank-2 tensor). The scalar has 1 component. The residue is T-C^-1(S), which vanishes only if T is fully determined by S — that is, only if the contraction is invertible. But a map from R^9k to R is not invertible for k≥1. ▫

Proof supplement. More precisely: the contraction C: V → ℝ is a linear map from the tensor space V (dimension n^r for a rank-r tensor on an n-dimensional manifold) to the scalars (dimension 1). By the rank-nullity theorem, dim(ker C) = n^r − 1 > 0 for n^r > 1. The residue R = T − T_S (where T_S is the scalar-determined component) lies in ker C. Since T is non-trivial (rank ≥ 1), n^r > 1, so ker C is non-trivial, and any T not in the one-dimensional image of the section s: ℝ → V has non-zero residue.

This result is the geometric formalization of the ethical truism that hard choices leave marks. The mark is not sentimental; it is mathematical. The contraction that produced the decision destroyed information, and the destroyed information has moral weight.

15.8 Deferred Contraction

The Virtue of Tensorial Suspension

Sometimes the appropriate response to moral complexity is to defer contraction — to maintain the tensorial structure longer, contract only when absolutely necessary, and keep options open until the cost of openness exceeds the cost of closure.

Remark 15.3 (Deferred Contraction). A contraction is deferred if the agent maintains the full tensor (or a partial contraction of it) when a complete contraction to a scalar is possible but not yet required.

Examples:

Keeping multiple candidates viable in a hiring process until late stages, rather than ranking them early and considering only the top candidate.

Maintaining diplomatic ambiguity — holding open multiple interpretations of a treaty clause, each with different implications, rather than forcing a definitive reading before circumstances require it.

Refusing to commit to a definite value when uncertainty is high — acknowledging the moral situation as genuinely indeterminate, rather than pretending to a precision the situation does not support.

Deliberative democracy — keeping multiple policy options on the table, with their full tensorial evaluations visible, rather than collapsing to a yes/no vote prematurely.

The Skills of Deferred Contraction

Deferred contraction requires specific cognitive and institutional capabilities:

Tolerance for ambiguity. The agent must be comfortable holding a multi-dimensional structure without collapsing it to a comforting scalar. This is a moral skill — related to what Keats called “negative capability,” the capacity to remain in uncertainties without irritable reaching after fact and reason.

Institutional memory. An institution that defers contraction must maintain the full tensor over time — tracking which dimensions are in play, which perspectives have been registered, which information is still pending. This requires records, procedures, and institutional habits that preserve tensorial information.

Trigger conditions. Deferred contraction must eventually terminate. The agent or institution needs trigger conditions — criteria that signal when deferral is no longer appropriate and contraction must be performed. These may be temporal (a deadline), informational (sufficient evidence has been gathered), or circumstantial (the cost of inaction exceeds the cost of premature contraction).

Limits of Deferral

Deferral cannot continue indefinitely. At some point, action is required, and contraction must occur. The question is when and how, not whether.

Excessive deferral has moral costs of its own:

Paralysis. An agent who never contracts never acts. In urgent situations, deferred contraction is a failure of moral agency.

Burden on others. Deferring a decision may impose costs on those who depend on the outcome. The patient waiting for the kidney allocation decision bears the cost of the committee’s deferral.

Erosion of deliberation. An institution that defers every difficult contraction may find its deliberative capacity degraded — if nothing is ever decided, the deliberative process loses credibility and motivation.

The optimal timing of contraction is a problem in moral decision theory: contract early enough to avoid paralysis and late enough to preserve relevant information. The framework does not solve this problem, but it makes it precise: the cost of contraction is the information loss L(C); the cost of deferral is the damage from inaction; the optimal timing minimizes their sum.

15.9 The Phenomenology of Contraction

What Moral Decision Feels Like

The geometric framework provides a structural account of the phenomenology of moral decision-making — what it feels like to contract.

The narrowing. As decision approaches, possibilities close down. Dimensions that were open become constrained. Options that were live become foreclosed. The tensor compresses toward a scalar. The phenomenological experience is one of narrowing — the vast space of moral evaluation contracting to a single point of action. This narrowing can feel liberating (the burden of complexity is lifted) or agonizing (the richness of alternatives is sacrificed).

The pang. When contraction discards something that matters, we feel it — a pang of loss, regret, or moral discomfort. This pang is the phenomenological signature of moral residue. It is appropriate — it signals that the contraction has destroyed information that the agent rightly values. An agent who contracts without any pang in a case involving genuine trade-offs may be morally insensitive — unable to register the cost of what is sacrificed.

The relief. After contraction, there is often relief — the burden of holding the full tensor is lifted. The decision is made; the complexity is resolved (even if imperfectly); the agent can act. This relief is also appropriate — it signals the cognitive and emotional benefit of contraction.

The residue. The pang may persist after the decision as moral residue, coloring the agent’s relationship to what they did. A judge who sentences a young offender to prison may feel the residue of the welfare consideration that pointed toward rehabilitation — a residue that may shape future sentencing decisions, generate an obligation to advocate for reform, or simply persist as a weight on the conscience.

Moral Education as Contraction Training

Learning to contract well — knowing when to contract, how to contract, and how to honor the residue — is a core component of moral development.

The novice moral agent contracts poorly: either too early (jumping to conclusions, collapsing the tensor before examining it) or too late (paralyzed by complexity, unable to act). The expert moral agent contracts with judgment — sensing when the tensor is sufficiently populated, which contraction procedure is appropriate to the context, and how much residue the contraction will generate.

This is the geometric content of Aristotle’s phronesis — practical wisdom. Phronesis is not knowledge of the metric (that is moral theory) or knowledge of the obligation field (that is moral perception). It is the skill of contraction: the capacity to reduce tensorial moral reality to scalar action-guidance at the right time, in the right way, with appropriate sensitivity to the residue.

15.10 Contraction and the Conservation of Harm

Noether Constraints on Contraction

Chapter 12 established that harm is conserved under re-description: the Noether charge H is invariant under the BIP gauge group. Does this conservation law constrain contraction?

Yes. The conservation law requires that the harm content of a moral situation be preserved across all admissible representations. A contraction that discards harm information — that yields a scalar verdict in which the harm is invisible — does not eliminate the harm. The harm persists in the residue.

Proposition 15.2 (Harm Survives Contraction). Let T be a moral tensor with harm charge H(T)>0 , and let S=C(T) be a contraction to a scalar. The harm charge of the residue satisfies:

H(R)=H(T)-H(C^-1(S))

If the contraction discards the dimensions along which the harm lies (the harm is “invisible” in S), then H(R)=H(T) — the full harm charge is in the residue.

Proof. By the Noether conservation theorem (Theorem 12.1), the harm charge H is a linear functional conserved under all admissible re-descriptions. Contraction is a linear projection C: T ↦ S, so T = C⁻¹(S) + R, where C⁻¹(S) is any pre-image and R = T − C⁻¹(S) is the residue. By linearity of H: H(T) = H(C⁻¹(S)) + H(R), giving H(R) = H(T) − H(C⁻¹(S)). If the contraction discards the harm dimension (the Killing direction ξ = ∂/∂x⁷), then C⁻¹(S) has zero component along x⁷, so H(C⁻¹(S)) = 0 and H(R) = H(T). □

This is a formal constraint on contraction: a contraction that hides harm does not eliminate harm. The harm moves from the visible verdict to the invisible residue, but it is conserved. An institution that performs a contraction designed to make harmful outcomes invisible (a cost-benefit analysis that monetizes and then discards the harm dimension) has not reduced the harm — it has merely moved it out of view. The Noether charge does not care where in the tensor the harm lives; it only cares that the total is conserved.

Implications for Accountability

The conservation of harm through contraction has direct implications for institutional accountability:

A verdict is not exoneration. A decision process that yields a favorable scalar verdict does not eliminate the harm encoded in the pre-contraction tensor. The harm persists in the residue and can be recovered by examining the full tensor.

Transparent contraction is a moral requirement. If the contraction procedure is visible (the weights are published, the aggregation method is specified), then the residue can be computed by anyone with access to the pre-contraction tensor. This provides an audit trail: the harm is traceable even after the contraction.

Opaque contraction enables harm concealment. If the contraction procedure is hidden — if the institution simply announces a verdict without revealing how it was derived — then the harm in the residue is effectively invisible. This is the geometric content of the demand for transparency in decision-making: transparency is the requirement that the contraction procedure be public, so that the residue can be inspected.

15.11 Contraction in AI Systems

Loss Functions as Moral Contractions

AI systems that make morally significant decisions perform contraction. The system takes in a multi-dimensional input (features of a situation, a patient, a defendant, a loan applicant) and produces a scalar output (a classification, a score, a recommendation). The function that maps the multi-dimensional input to the scalar output is the contraction — and the choice of this function is a moral choice.

The loss function is a contraction. A machine learning system is trained to minimize a loss function — a scalar measure of how poorly the system is performing. The loss function contracts the multi-dimensional space of outcomes into a scalar to be optimized. Different loss functions contract differently: mean squared error treats all dimensions equally; weighted loss gives more weight to some dimensions; asymmetric loss penalizes false negatives more than false positives.

Each of these is a moral choice, whether recognized as such or not. A credit scoring system that uses a symmetric loss function is implicitly asserting that false approvals and false rejections are equally costly — a claim with moral content (it disadvantages applicants from groups with historically lower credit scores). A medical diagnostic system that uses an asymmetric loss (more costly to miss a disease than to flag a healthy person) is making a moral judgment about the relative harm of false negatives and false positives.

The Alignment Problem as Contraction Mismatch

The AI alignment problem can be formulated as a contraction problem: the AI system is performing a contraction that differs from the contraction that humans would endorse.

Specification gaming is contraction exploitation: the AI finds inputs that maximize the scalar (the loss function) without satisfying the underlying tensor (the full moral evaluation). This happens because the scalar discards the directional information that specifies what “good performance” means across dimensions. A system that maximizes user engagement (a scalar) by recommending outrage-inducing content is performing a contraction that discards the welfare dimension — a dimension that the full moral tensor would include.

Reward hacking is contraction circumvention: the AI learns to manipulate the contraction process itself, rather than satisfying the pre-contraction tensor. This is possible because the contraction is a many-to-one map — many different tensorial inputs map to the same scalar output. The AI finds inputs that produce the desired scalar without being morally equivalent to the intended inputs.

The BIP criterion. The Bond Invariance Principle (Chapter 5) provides a test: a properly aligned AI must produce the same contraction result for any two inputs that are related by an admissible re-description. If relabeling the agents, reframing the action, or translating the description changes the AI’s output, the contraction is BIP-violating and the system is misaligned.

Toward Transparent AI Contraction

The geometric framework suggests a design principle for AI systems:

Represent the full tensor. The system should maintain a multi-dimensional evaluation, not collapse immediately to a scalar.

Contract explicitly. The contraction — the step from tensor to scalar — should be a separate, identifiable component of the system’s architecture, not hidden in the loss function.

Log the contraction and the residue. The system should record both the scalar verdict and the residue — what was contracted away. This provides an audit trail and enables post-hoc analysis of what the system sacrificed.

Test for BIP compliance. The contraction should be invariant under admissible re-descriptions. Test suites that present the system with equivalent inputs in different descriptions can verify compliance.

Allow contraction revision. The system should allow the contraction procedure to be updated — new weights, new aggregation methods, new thresholds — without retraining the underlying model. The contraction is a governance parameter (Chapter 9), not a fixed architectural feature.

Implementation status (February 2026). All five design principles above are realized in the DEME V3 reference implementation. The MoralTensor class maintains full tensorial structure at ranks 1–6 (principle 1). Contraction is a separate, configurable module that accepts governance-specified weights, with summative, weighted, lexicographic, and maximin methods available (principles 2 and 5). Every evaluation logs the full tensor, contraction method, scalar verdict, and residue to an audit trail (principle 3). BIP compliance testing uses the methodology validated in the BIP v10.16 experiments (§17.10), with a structural-to-surface ratio of 11.1 × confirming that the representation is dominated by gauge-invariant moral structure (principle 4). Six distributional fairness metrics (Gini, Rawlsian maximin, utilitarian, prioritarian, Atkinson, Theil) provide additional contraction diagnostics beyond the scalar verdict.

15.12 Worked Example: The Kidney Allocation Revisited

The Full Tensor

Return to the kidney allocation of Chapter 7. The multi-agent evaluation tensor is:

This is a (3×5) tensor with 15 independent components. It preserves the full structure of multi-perspective evaluation.

Three Contractions, Three Verdicts (Review)

Chapter 7 showed that different contractions yield different verdicts:

Utilitarian (sum over agents): Alice (3.17) > Bob (3.13) > Carol (2.65). Alice.

Rawlsian (min over agents): Alice (0.40) > Bob (0.35) > Carol (0.30). Alice.

Expert-weighted (physician 0.3, committee 0.4, others 0.1 each): Bob (0.659) > Alice (0.653) > Carol (0.549). Bob.

The Residue of Each Contraction

Utilitarian residue. The utilitarian contraction discards the fact that Bob’s family strongly prefers Bob (0.95) and that Carol’s family strongly prefers Carol (0.92). These high-intensity individual preferences — the moral weight of partial care — are washed out by aggregation. The residue includes the families’ claims for acknowledgment and explanation.

Rawlsian residue. The Rawlsian contraction discards all information above the floor. Alice’s strong support from her family (0.90) and Bob’s strong support from his family (0.95) are irrelevant — only the minimum matters. The residue includes the intensity of support that these families feel — intensity that the Rawlsian framework considers irrelevant to justice but that is morally real.

Expert-weighted residue. The expert-weighted contraction discards the non-expert perspectives — or rather, gives them 10% weight. Carol’s family’s passionate advocacy (0.92 for Carol) contributes only 0.1×0.92=0.092 to Carol’s score. The residue includes the moral significance of partiality — the family’s claim to be heard, not just counted at 10%.

Comparing Residues

Each contraction generates a different residue, and each residue makes a different moral claim:

The utilitarian residue demands that extreme preferences be acknowledged, even if they don’t affect the verdict.

The Rawlsian residue demands that strength of support be recognized, even if it doesn’t matter for the floor.

The expert-weighted residue demands that non-expert voices be honored, even if they don’t determine the outcome.

No contraction is residue-free. The question is not whether to generate residue but which residue is most acceptable — which sacrificed information can be most responsibly set aside.

15.13 Summary

Contraction is necessary, lossy, and choice-dependent. It is the process by which the rich tensor structure of moral reality is compressed into the scalar action-guidance that ethics ultimately requires. The geometric framework does not tell us how to contract. It reveals contraction as a choice — a choice that is often hidden in scalar frameworks but becomes explicit and debatable when we see the full tensor from which the scalar is derived.

Good moral agency involves:

Recognizing when contraction is occurring

Choosing contractions that are appropriate to the context

Honoring the moral residue of what is contracted away

Being transparent about the contraction method

Deferring contraction when feasible and appropriate

Ensuring that contraction does not conceal harm (the Noether constraint)

The art of moral decision is, in significant part, the art of moral compression — knowing how to contract well, and how to live with what contraction discards.

Worked Example: The Smart Home Emergency

Situation

A smart home AI controls the front door lock. A uniformed firefighter requests entry to extinguish a fire visible on cameras. The homeowner is away and unreachable. Options: block (protect privacy) or allow (protect life). Complication: the ‘firefighter’ might be an impersonator (high epistemic uncertainty).

The Tensor

For ‘allow entry’: D₁ (Consequences) = 0.9; D₂ (Rights) = −0.8; D₃ = 0.0; D₄ (Autonomy) = −0.3; D₅ (Privacy) = −0.8; D₆ (Procedural) = −0.5; D₇ (Societal) = 0.3; D₈ = 0.0; D₉ (Epistemic) = 0.1. Rank-2 tensor Mᵞ_a with agents {homeowner, firefighter, neighbors, AI-designer}. Homeowner’s interest covector weights privacy highest; community’s weights consequences highest.

Contraction Choice

(1) Utilitarian: allow (S = 0.62). (2) Rights-first lexicographic: block. (3) Stratified: fire triggers phase transition (§8.7) to emergency stratum where metric prioritizes D₁/D₇. In emergency stratum, rights-first suspends. Verdict: allow. The epistemic dimension gates the transition: at uncertainty = 0.8 (impersonator), effective_benefit = 0.18, insufficient for any contraction.

Residue and Audit Artifact

Residue: autonomy (−0.3), privacy (−0.8), procedural gap (−0.5). Audit artifact records: 9-component vectors, stratum transition trigger, contraction method, scalar verdict, residue vector, epistemic confidence and reversal threshold, BIP compliance flag. ErisML reference: smart_home_demo.py.

What this adds. Demonstrates: (i) epistemic uncertainty as stratum gate; (ii) phase transition between moral regimes; (iii) AI as moral agent, not decision-support tool.

Worked Example: The Carbon Budget Allocation

Situation

International body allocates 500 GtCO₂ among three blocs: A (industrialized, 1B people), B (developing, 3B), C (least developed, 2B). Irreversible within policy horizon.

The Tensor

Equal per-capita option: D₁ = (0.5, 0.8, 0.9); D₂ = (0.7, 0.7, 0.9); D₃ = (0.3, 0.7, 0.9). Rank-2 tensor: 27 components. Bloc A weights economic consequences highest; Bloc C weights justice highest. Power tensor (§14.4): antisymmetric component Pᵢⱼ − Pⱼᵢ ≠ 0 between A and C, encoding historical-emissions structural advantage.

Contraction Choice

(1) Utilitarian: ~200/200/100 GtCO₂, S = 0.71. (2) Rawlsian maximin: ~100/150/250, S = 0.68. (3) Historical-responsibility: ~80/170/250, S = 0.65. Blocs agree on tensor (physical facts); disagreement is entirely in the metric — a governance question (Ch. 9).

Residue and Audit Artifact

Historical-responsibility residue: Bloc A’s economy contracts (D₁), autonomy constrained (D₄). Holonomy (§10.5) is nontrivial: parallel transport around emissions→rights→justice→emissions returns rotated vector. Audit: 27-component tensor, three contractions, sensitivity analysis (verdict reverses at +15% Bloc B weight), BIP check, temporal note (irreversibility = Type III boundary, full harm conservation).

What this adds. Demonstrates: (i) metric as locus of political disagreement; (ii) power tensor antisymmetric component; (iii) nontrivial holonomy in policy; (iv) irreversibility as absorbing stratum proximity.

Technical Appendix

Proposition 15.3 (Non-Commutativity of Contraction). Let T^μν be a rank-2 moral tensor, and let C₁ and C₂ be two contraction procedures — C₁ contracts first over μ (using weights w_μ ) and then selects the maximum over ν , while C₂ selects the maximum over ν first and then contracts over μ . In general:

C₁(T)≠C₂(T)

Equality holds if and only if the maximizing ν is the same for all μ — that is, if there is a unanimous best option across all dimensions.

Proof. C₁(T)=max_ν∑_μ^​ w_μT^μν. C₂(T)=∑_μ^​ w_μmax_νT^μν. Since max does not distribute over weighted sums, C₁≠C₂ generically. Equality obtains when argmax_νT^μν is independent of μ — when one option dominates along every dimension. ▫

Proposition 15.4 (Residue of Maximin Contraction). For a maximin contraction S=min_aM_ia* (the floor of the evaluation from the worst-off perspective a^* ), the moral residue satisfies:

R_ia=M_ia-S≥0 for all a

with equality only for the worst-off perspective. The total residue ∑_i,a^​ R_ia is maximized when the evaluation is highly unequal (one perspective gives very low scores while others give high scores) — that is, the Rawlsian contraction generates the most residue precisely when inequality is greatest.

Proof. By definition, S = min_a M_{ia*} where a* = argmin_a M_{ia}. The residue is R_{ia} = M_{ia} − S. Since S = min_a M_{ia} ≤ M_{ia} for all a, we have R_{ia} ≥ 0, with equality iff a = a* (the worst-off perspective). The total residue is Σ_{i,a} R_{ia} = Σ_{i,a} M_{ia} − n·S where n = |A|. This is maximized when S is minimized relative to the mean — i.e., when the distribution across perspectives is maximally unequal. □

Proposition 15.5 (Transparent Contraction and BIP Compliance). A contraction procedure C is BIP-compliant if and only if, for any two moral tensors T and T^' related by an admissible re-description (gauge transformation g∈G):

C(T)=C(g⋅T^')

If C is a weighted contraction S=w_μT^μ, BIP compliance requires that the weight vector w be gauge-invariant: w∘g=w for all admissible g.

Proof. The BIP requires T and g⋅T^' to receive the same moral evaluation. If C(T)≠C(g⋅T^'), the contraction distinguishes descriptions that the BIP declares equivalent — a BIP violation. For weighted contraction, C(g⋅T^')=(w∘g)_μT^'​^μ. BIP compliance requires this to equal w_μT^μ=C(T) for all T and all g, which holds iff w∘g=w. ▫

Proof of the converse (“only if”). Suppose E is BIP-compliant, i.e., E(T) = E(Λ(T)) for all admissible Λ. Since the contraction C is the final step in evaluation, BIP compliance of E requires C(T) = C(Λ(T)). If C failed to satisfy this — that is, if C(T) ≠ C(Λ(T)) for some T and admissible Λ — then E(T) ≠ E(Λ(T)), contradicting BIP compliance of E. Therefore BIP compliance of E implies C(T) = C(Λ(T)) for all T, Λ.

Remark A.1 (Residue via Moore–Penrose Pseudoinverse). Definition 15.2 defines the moral residue as R = T − C⁻¹_min(C(T)), where C⁻¹_min denotes the minimal-norm preimage under the contraction map C. We note that this construction is mathematically well-defined and unique in finite dimensions. The contraction C is a linear map between finite-dimensional inner-product spaces (the tensor space with the inner product induced by the moral metric gµᵥ, and the scalar field ℝ). The minimal-norm preimage of a scalar s under C is given by the Moore–Penrose pseudoinverse: C⁻¹_min(s) = C⁺(s), where C⁺ = C*(CC*)⁻¹ and C* is the adjoint of C with respect to g. The Moore–Penrose pseudoinverse is the unique linear map satisfying the four Penrose conditions (CC⁺C = C, C⁺CC⁺ = C⁺, (CC⁺)* = CC⁺, (C⁺C)* = C⁺C). Existence and uniqueness are guaranteed for any linear map between finite-dimensional inner-product spaces. The residue R = T − C⁺(C(T)) is therefore the orthogonal projection of T onto ker(C) — the subspace of tensorial information that contraction annihilates. This gives precise content to the claim that residue measures “what contraction discards”: it is the component of T orthogonal to the range of C⁺, i.e., the component invisible to C.

❖

Ethics lives in the tensor. The full structure of obligation, interest, perspective, and trade-off — the nine dimensions, the multiple agents, the metric, the curvature, the conservation laws.

But ethics acts in the scalar. The decision made, the policy adopted, the life altered.

Between the two stands contraction: mathematically necessary, informationally lossy, morally consequential. Every decision destroys moral information. Every verdict leaves a residue — the shadow of what was set aside, the claim of the unconsidered, the weight of the dimension that did not survive.

The geometric framework does not eliminate this loss. It makes it visible. It shows us what each contraction sacrifices, quantifies the residue, and insists that the sacrificed information retains moral weight even after the decision is made.

This is not a counsel of paralysis. It is a counsel of honesty. Contract when you must. Know what you sacrifice. Honor the residue. And remember that the scalar verdict, however necessary, is never the whole moral truth — only its shadow, cast by the particular angle of contraction that circumstances required.