Appendix A: Related Work and Differentiation

This appendix positions geometric ethics relative to the nearest prior art in five domains: mathematical ethics, AI alignment, formal verification, gauge theory outside physics, and invariance in decision theory. For each domain, we identify what the prior approach captures, what it misses, and what geometric ethics adds. A summary comparison table appears at the end.

A.1 Mathematical Ethics

Utilitarian calculus (Bentham 1789; Mill 1863; Harsanyi 1955) provides a scalar aggregation of welfare: the right action maximizes the sum (or expected value) of individual utilities. This captures the commensurability intuition — that moral options can be ranked — but achieves it by discarding all tensor structure. In geometric ethics, utilitarianism corresponds to a specific contraction procedure (Chapter 15) applied to the full moral tensor. The contraction loss theorem (Proposition 15.1) proves that this reduction is always lossy: the residue R is nonzero for any non-trivial tensor. Utilitarianism is not wrong; it is incomplete in a precisely quantifiable sense.

Social choice theory (Arrow 1951; Sen 1970, 1999) formalizes collective preference aggregation and proves impossibility results (Arrow’s theorem, Sen’s liberal paradox) that constrain any aggregation rule. These results operate on ordinal rankings and do not engage with the geometric structure of the option space. Geometric ethics provides the missing geometry: the moral manifold M, the metric g, and parallel transport (Chapter 10) allow meaningful comparison of moral evaluations across different situations — precisely the kind of cross-situational comparison that social choice theory treats as undefined.

Rawlsian maximin (Rawls 1971) prioritizes the worst-off member of society, implementing a lexicographic ordering over welfare levels. In geometric ethics, this corresponds to a specific metric choice: one in which the lowest-welfare dimension has infinite weight relative to all others — a degenerate metric that creates a lexicographic stratum (Definition 5.4). The framework thus embeds Rawlsian justice as a special case while allowing the metric to vary continuously between utilitarian and lexicographic extremes.

Contractualism (Scanlon 1998) grounds moral principles in what no one could reasonably reject. This captures the structural intuition that moral evaluation should be perspective-invariant, but provides no formal mechanism for testing invariance. The Bond Invariance Principle (Axiom 5.1) is the geometric realization of this intuition: legitimate moral evaluations are gauge-invariant quantities, and the D₄ × U(1) gauge group (Theorem 12.3) provides the explicit symmetry structure that contractualism gestures at.

A.2 AI Alignment and Value Learning

Reinforcement learning from human feedback (RLHF; Christiano et al. 2017; Ouyang et al. 2022) trains AI systems to align with human preferences by optimizing a learned reward model. RLHF captures empirical preferences but provides no structural guarantee against reward hacking, distributional shift, or adversarial manipulation. Geometric ethics offers structural containment: the No Escape Theorem (Theorem 18.1) proves that norm compliance cannot be circumvented by any agent architecture operating within the DEME framework, because the norm check is upstream of action execution — the forbidden action is structurally impossible, not merely penalized.

Constitutional AI (Bai et al. 2022) equips language models with explicit rules and trains them to self-critique against those rules. This achieves rule compliance but lacks formal verification: there is no proof that the trained model actually follows the constitution in all cases. The DEME architecture (Chapter 19) provides what Constitutional AI lacks: a four-layer evaluation pipeline with tiered governance, cryptographically bound audit artifacts, and a formally verified veto mechanism. The Bond Index (Chapter 17) provides the quantitative compliance metric that Constitutional AI does not have.

Inverse reward design (Hadfield-Menell et al. 2017) and debate protocols (Irving et al. 2018) address alignment through preference inference and adversarial deliberation, respectively. Both assume that the correct moral evaluation exists implicitly in human feedback and can be extracted through clever querying. Geometric ethics takes the opposite stance: moral evaluation has explicit geometric structure (tensor rank, stratification, curvature) that is not reducible to preference data. The framework provides the formal language in which alignment guarantees can be stated and verified, rather than hoped for.

More broadly, the AI safety literature (Amodei et al. 2016; Hendrycks et al. 2021) identifies concrete problems — side effects, reward gaming, scalable oversight — but typically addresses them piecemeal. Geometric ethics provides a unified mathematical framework in which these problems have geometric interpretations: side effects are curvature (Chapter 10), reward gaming is gauge non-invariance (detectable via Bond Index), and scalable oversight is the audit layer (Chapter 19).

A.3 Formal Verification and Deontic Logic

Standard deontic logic (SDL; von Wright 1951) formalizes obligation, permission, and prohibition using modal operators. SDL is notorious for its paradoxes: the gentle murderer paradox, the contrary-to-duty paradox, and Ross’s paradox all arise from the inability of modal operators to handle exception structure. Geometric ethics replaces modal operators with Whitney stratification (Chapter 8): exceptions are not logical anomalies but stratum boundaries, and the semantic gate formalism (Definition 8.8) handles the transition between normal and exceptional moral regimes without paradox.

Input/output logic (Makinson and van der Torre 2000) and normware (Governatori 2015) extend deontic logic with defeasibility and computational implementation, respectively. These approaches handle exceptions better than SDL but remain within the propositional framework. Geometric ethics operates on a continuous manifold with metric, connection, and curvature — providing the infinitesimal structure that logic-based approaches lack. The satisfaction function S (Definition 6.3) is a smooth scalar field, not a truth value, enabling gradient-based moral reasoning that propositional approaches cannot express.

Runtime verification for normative systems (Alechina et al. 2012; Dennis et al. 2016) monitors agent behavior against formal specifications. Geometric ethics extends this with the BIP compliance methodology: rather than checking individual actions against rules, the Bond Index measures the system’s structural invariance under the full transformation suite. A Bd = 0 certificate is stronger than rule-by-rule compliance because it guarantees that no redescription of the input can change the evaluation — including redescriptions the verifier has not anticipated.

A.4 Gauge Theory Outside Physics

Information geometry (Amari 1985; Amari and Nagaoka 2000) applies Riemannian geometry to statistical manifolds, with the Fisher information metric playing the role of the moral metric g. The parallel is illuminating but limited: information geometry has no analog of stratification, no gauge group, and no normative content. Geometric ethics borrows the differential-geometric toolkit but applies it to a fundamentally different domain — one with discrete symmetries (D₄) alongside continuous ones (U(1)_H), and with absorbing strata that have no statistical analog.

Gauge theory in economics (Malaney and Weinstein 2014) applies fiber bundle structure to index number theory, showing that price indices are gauge-dependent quantities. This is the closest precedent to our approach: the insight that “what looks like a measurement is actually a gauge choice” applies directly to moral evaluation. However, Malaney-Weinstein use U(1) gauge symmetry only; geometric ethics introduces the non-abelian D₄ component, which captures the discrete symmetry structure specific to deontic relations.

Fiber bundles appear in robotics (Murray, Li, and Sastry 1994) and in computational anatomy (Younes 2010), where they model configuration spaces and shape spaces, respectively. These are structural parallels — the moral manifold M is, mathematically, the same kind of object — but the ethical semantics (strata as moral regimes, curvature as moral path-dependence, holonomy as accumulated moral change) are original contributions of geometric ethics.

A.5 Invariance in Decision Theory

Independence of irrelevant alternatives (Arrow 1951) requires that the ranking of two options not change when a third option is added or removed. Description invariance (Tversky and Kahneman 1981) requires that preferences not change under logically equivalent redescriptions. Framing effects (Kahneman 2011) are the empirical violations of description invariance. All three point to the same insight: legitimate evaluation should be invariant under certain transformations.

Geometric ethics provides the formal mechanism that these principles lack. The Bond Invariance Principle (Axiom 5.1) is not merely a normative requirement but a gauge symmetry with a specific group structure (D₄ × U(1)_H). The BIP testing methodology (Chapter 18.7) operationalizes invariance testing as a concrete computational procedure: apply the transformation suite, check that the verdict is unchanged, and report the Bond Index. Framing effects are not anomalies but gauge non-invariances, and the Bond Index quantifies their magnitude.

The closest precedent is Sunstein’s work on commensurability (Sunstein 1994), which argues that some moral dimensions are genuinely incommensurable. In geometric ethics, incommensurability is a metric property: dimensions μ and ν are incommensurable when g_{μν} = 0 (the off-diagonal metric component vanishes) and no coordinate change can make them commensurable. This is a degenerate metric (Definition 6.2), not a logical impossibility.

A.6 Comparison Table

The following table summarizes the relationship between geometric ethics and the nearest prior art across the five domains reviewed above.

The consistent pattern across all domains is the same: prior approaches capture important moral intuitions but lack the geometric structure needed to formalize them. Geometric ethics provides that structure — not as a replacement for prior work, but as a mathematical framework in which the insights of prior work can be stated precisely, compared rigorously, and tested computationally.