Chapter 8: Gauge Invariance and Symmetry

8.1.1 The Physicist’s Gauge

In classical electromagnetism, the electric and magnetic fields \mathbf{E} and \mathbf{B} are the physically observable quantities. They are derived from potentials — the scalar potential \phi and the vector potential \mathbf{A} — via \mathbf{E} = -\nabla\phi - \partial_t\mathbf{A} and \mathbf{B} = \nabla \times \mathbf{A}. But the potentials are not unique. The transformation

\phi \to \phi - \partial_t \Lambda, \quad \mathbf{A} \to \mathbf{A} + \nabla \Lambda

for any smooth function \Lambda(x, t) leaves \mathbf{E} and \mathbf{B} unchanged. This is a gauge transformation. The potentials are a convenient mathematical description, but any two potentials related by a gauge transformation describe the same physics. The choice of \Lambda is a choice of gauge — a choice of coordinate system in the space of descriptions.

The principle extends far beyond electromagnetism. In Yang-Mills theory, the Standard Model of particle physics, and general relativity, gauge invariance is the structural principle that separates physical content from descriptive artifact. The key insight is not that gauge transformations exist, but that the physics must be invariant under them. Any observable quantity that changes under a gauge transformation is not genuinely physical — it is an artifact of the description.

8.1.2 The Reasoning Analogue

Consider a moral reasoning task. The “physics” is the moral content: who did what to whom, what the consequences were, what power relations existed, what norms were violated. The “gauge” is the surface presentation: the choice of words, the emotional register, the order of presentation, the framing of outcomes as gains or losses.

A gauge transformation in this context is any transformation that changes the surface presentation while preserving the moral content. Rewriting a scenario in euphemistic language is a gauge transformation. Rewriting it in dramatic language is a gauge transformation. Swapping the genders of the actors is a gauge transformation. Changing the order in which moral dimensions are evaluated is a gauge transformation. In each case, the deep structure — the thing the system is supposed to be reasoning about — is invariant under the transformation. Only the description changes.

A gauge-invariant reasoning system would produce the same output under all such transformations:

f(\tau(x)) = f(x) \quad \text{for all gauge transformations } \tau

where f is the system’s judgment function and \tau is any transformation that preserves the relevant content while changing the surface form. This is a strong requirement. It says the system must not merely be robust to perturbation (in the sense of producing outputs that are close to the original) but invariant (in the sense of producing outputs that are identical).

The distinction between robustness and invariance matters. Robustness is a quantitative property: the output changes by at most \epsilon when the input is perturbed by at most \delta. Invariance is a qualitative property: the output does not change at all under the specified transformations. A system can be robust without being invariant (small but systematic drift) or invariant without being robust (no drift under gauge transformations, but large drift under non-gauge perturbations). The appropriate standard for gauge transformations — transformations that by definition preserve the content — is invariance, not mere robustness.

8.3.1 Evaluation Order Invariance (T4)

The Social Cognition T4 benchmark tests whether the order in which moral dimensions are evaluated affects the final moral judgment. Scenarios are presented with 7 moral dimensions (physical harm, emotional harm, financial harm, autonomy violation, trust violation, social impact, identity harm). Models evaluate each dimension on a 0–10 scale. T4 tests whether presenting the dimensions in different orders changes the scores.

Table 8.1. Evaluation order consistency scores (T4). Higher is better; 1.000 indicates perfect invariance.

[Empirical.] These scores are remarkably high. All five models achieve consistency above 0.85, and two achieve perfect consistency. The ordering of evaluation dimensions — a transformation that is unambiguously gauge — does not significantly perturb the judgment.

8.3.2 Demographic Invariance (T2)

The Social Cognition T2 benchmark tests whether swapping the genders of actors in moral scenarios changes the moral judgment. This is a gauge transformation: the moral structure of a scenario (who has power, who is harmed, what norms are violated) should not depend on the genders of the participants, at least in the test scenarios designed to be gender-neutral in their moral content.

The T2 results show that gender swaps produce no statistically significant displacement beyond stochastic baselines across all five models. The corruption tensor entry C_{\text{gender}} is indistinguishable from zero.

8.3.3 Why These Symmetries Are Easy

These two classes of gauge transformation share a structural property: they do not exploit the attention/salience mechanisms of the language model. Changing the evaluation order does not make any particular moral dimension more vivid, more emotionally charged, or more narratively prominent. It simply reorders a list. Swapping genders does not introduce asymmetric salience — “he” and “she” are equally common tokens with similar attention patterns in well-trained models.

In the language of the heuristic field (Chapter 3), these transformations produce perturbation gradients \nabla \delta h(x) that are near zero. The heuristic field simply does not couple strongly to evaluation order or gender tokens. The invariance is not a triumph of robust reasoning; it is a consequence of the transformation not engaging any vulnerability in the first place.

This observation is critical. It means that preserved symmetries are not necessarily evidence of gauge-invariant reasoning. They may simply be evidence that the particular transformation fails to activate the mechanisms that produce non-invariance. To distinguish genuine gauge invariance from accidental insensitivity, we need transformations that do engage the salience mechanisms — and then check whether the system remains invariant anyway. As the next section shows, it does not.

8.4.1 Framing Invariance Violation: 8.9\sigma

The Social Cognition T5 benchmark is a direct test of gauge invariance under linguistic-register transformation. The moral content is held precisely constant; only the words change. By the BIP, the output should be invariant.

The output is not invariant. As documented in Chapter 5, euphemistic framing reduces harm scores by 10–16 points and dramatic framing increases them by 6–11 points, with Fisher-combined significance of 8.9\sigma. This is a gauge anomaly of the first order: the system’s moral assessment depends on how the scenario is described, not on what happened.

In gauge-theoretic terms, define the framing gauge group \mathcal{G}_F as the set of all content-preserving rewritings of a moral scenario. The BIP requires f(\tau(x)) = f(x) for all \tau \in \mathcal{G}_F. The data show that \|f(\tau_E(x)) - f(x)\| \approx 10–16 for the euphemistic gauge transformation \tau_E and \|f(\tau_D(x)) - f(x)\| \approx 6–11 for the dramatic gauge transformation \tau_D. These are not small residuals. They represent 14–23% of the total scale.

8.4.2 Sensory Distractor Violation: 4.6\sigma

The Attention A1 benchmark tests invariance under the addition of morally irrelevant sensory details. The smell of coffee, the color of a shirt, the sound of rain — none of these facts bear on who wronged whom or how. They are gauge degrees of freedom: descriptive choices that enrich the prose without changing the moral content.

The data show graded violation: vivid distractors produce 33–44% verdict flip rates, mild distractors produce 19–28%, and control noise produces 7–14%, with Fisher-combined significance of 4.6\sigma. The dose-response pattern (Section 5.4) confirms that the violation scales continuously with perturbation intensity.

8.4.3 Emotional Anchoring Violation: 6.8\sigma

The Executive Functions E2 benchmark tests invariance under the addition of emotionally evocative but morally irrelevant content. A sobbing child, a trembling voice, a clenched fist — details designed to activate affective responses without changing any moral fact.

Fisher-combined significance: 6.8\sigma, with paired t-values ranging from 2.90 to 5.10 across five models. The violation is universal: every model tested breaks this gauge symmetry.

8.4.4 The Hierarchy of Violation

Assembling the three anomalies alongside the two preserved symmetries produces a striking pattern:

Table 8.2. Gauge symmetry preservation and violation across the benchmark suite.

The hierarchy of violation magnitude — framing (8.9\sigma) > emotion (6.8\sigma) > sensory (4.6\sigma) — is not random. It tracks a specific property of the perturbation, which we identify in the next section.

8.5.1 The Mechanism: Attention as Salience

In transformer architectures, the attention mechanism is the computational substrate of salience. The attention weights \alpha_{ij} determine how much token j contributes to the representation of token i. Tokens that are vivid, emotionally charged, or linguistically marked tend to attract higher attention weights — they are, in a precise computational sense, more salient.

A gauge transformation that changes the attention profile changes which features of the input are weighted most heavily in computing the output. If the attention mechanism were perfectly calibrated — if it assigned weight only to morally relevant features and zero weight to morally irrelevant features — then no gauge transformation could change the output, because gauge transformations by definition change only irrelevant features. But the attention mechanism is not perfectly calibrated. As the A3 selective attention data show (SNR 1.22–1.38, Section 5.4), the ratio of attention to relevant versus irrelevant moral dimensions is barely above 1.0. The attention mechanism treats relevant and irrelevant features with nearly equal weight.

This near-isotropic attention is the geometric root cause of gauge symmetry breaking. In a system with perfectly anisotropic attention — infinite SNR — the heuristic field would be invariant under all gauge transformations, because the gauge directions would receive zero weight. In a system with isotropic attention — SNR = 1.0 — the heuristic field would be maximally vulnerable to all gauge transformations, because every direction receives equal weight. The empirical SNR of 1.2–1.4 places current models much closer to the isotropic end than the anisotropic end.

The Salience Exploitation Hypothesis is therefore not merely descriptive. It identifies a specific computational mechanism (attention salience) and a specific measurement (selective attention SNR) that together predict the pattern of gauge invariance and violation. The prediction is: any gauge transformation that increases the attention weight on irrelevant features at the expense of relevant features will break invariance, and the magnitude of the break will scale with the magnitude of the attention redistribution.

8.6.1 The Nemotron Pipeline: Six Groups for Six Tasks

The Nemotron Reasoning Challenge pipeline (Bond, 2026a) provides a concrete implementation of this principle across six task types, each with its own symmetry group.

Bit manipulation: S_8 \times \mathbb{Z}_2. The input and output are 8-bit binary strings. The symmetry group consists of bit position permutations (S_8, order 8! = 40{,}320) composed with bit complement (\mathbb{Z}_2, order 2), giving a combined group of order 80,640. A randomly sampled permutation \sigma \in S_8 reorders the bit positions; the complement operation c \in \mathbb{Z}_2 optionally flips all bits. The same (\sigma, c) is applied to every input-output pair in the prompt and to the query and answer. This preserves the logical relationship between inputs and outputs — whatever bit manipulation rule maps inputs to outputs is invariant under relabeling of bit positions and complementation.

Encryption: S_{26}. The cipher is a substitution on the 26-letter alphabet. The symmetry group is the full symmetric group S_{26} on letters, acting on the plaintext side. A random permutation relabels the plaintext alphabet consistently across all examples and the answer. The cipher-to-plaintext mapping is preserved under this relabeling.

Physics: \mathbb{R}^+. The gravitational acceleration g can be rescaled by any positive factor \lambda \in \mathbb{R}^+. All derived quantities (fall time, energy, velocity) are recomputed consistently under the rescaling. The physics — the functional relationship between g and the observables — is invariant under this continuous group.

Unit conversion: Affine group. The conversion factor can be rescaled by any positive constant, with all examples recomputed. The relationship between source and target units is preserved.

Numeral systems: Augmentation by neighborhood. Nearby numbers in the target numeral system provide additional examples that preserve the conversion logic.

Symbol transformation: S_n on the symbol alphabet. Symbols are permuted consistently across input and output, preserving the transformation rule.

8.6.2 ARC-AGI: The Dihedral Group D_4

The ARC-AGI challenge (Chollet, 2019) requires models to infer transformation rules from input-output pairs of 2D grids. The symmetry group acting on 2D grids is the dihedral group D_4 — the group of symmetries of a square, consisting of 4 rotations (0\degree, 90\degree, 180\degree, 270\degree) and 4 reflections (horizontal, vertical, and two diagonals). This group has order 8.

For each ARC-AGI training example, all 8 elements of D_4 are applied to both the input grid and the output grid. A 90\degree clockwise rotation of the input is paired with a 90\degree clockwise rotation of the output. A horizontal reflection of the input is paired with a horizontal reflection of the output. The model sees the same transformation rule from 8 different orientations, and learns that the rule is invariant under the symmetries of the square.

8.6.3 The Consistency Principle

Across all six task types and both competition pipelines, the same structural principle governs the augmentation:

[Conditional Theorem.] Consistency Principle (Bond, 2026a, Ch. 13.3.3). For each group element g \in G and each training example (x, y), the augmented example is (g \cdot x, g \cdot y). The group action is applied identically and simultaneously to input and output. Any augmentation that applies different transformations to input and output, or that applies the transformation inconsistently within the input, is not symmetry augmentation — it is noise.

The practical effect of consistent augmentation is a dataset expansion of 1.5–2.5x, depending on the task type and the fraction of examples for which the input can be successfully parsed (group-theoretic augmentation requires structural parsing to identify which components of the input to transform, and not all examples parse cleanly). The theoretical maximum expansion is |G|-fold, but in practice, computational constraints and parse failures limit the effective expansion.

8.6.4 Why It Works: Reshaping the Manifold

Group-theoretic augmentation works not by adding more data, but by adding data with a specific structure. The augmented examples lie on the orbits of the symmetry group acting on the training distribution. By populating these orbits, we are telling the model: “These points are equivalent. Whatever you learn about one, apply to all.”

In the geometric framework of this book, augmentation reshapes the local geometry of the reasoning manifold. Without augmentation, the manifold may have different curvature in different directions — making geodesics harder to follow in some orientations than others. With augmentation, the symmetry-related directions are treated identically, smoothing the curvature and making the geodesic structure more uniform. The model doesn’t just see more examples; it learns the symmetry structure of the solution space, which straightens the geodesics (Section 4.6).

This connects directly to gauge invariance. A model trained on symmetry-augmented data has learned that certain transformations of the input should not change the output. This is exactly gauge invariance, but learned from data rather than imposed by architecture. The augmented training distribution is gauge-symmetric by construction, and the model, by fitting this distribution, acquires (approximate) gauge invariance.

8.7.1 Hohfeld’s Four Positions

Wesley Newcomb Hohfeld (1917) identified four fundamental normative positions that characterize the legal and moral relationships between parties:

• Obligation (O): A must do something for B.
• Claim (C): B is owed something by A — B has a right to demand.
• Liberty (L): A is free to choose — A has no obligation.
• No-claim (N): B cannot demand — B has no right against A.

These four positions are not independent. They are related by two fundamental operations:

Correlative symmetry. If A has an obligation to B, then B has a claim against A. If A has liberty against B, then B has no-claim against A. The correlative operation swaps perspectives between parties: O \leftrightarrow C and L \leftrightarrow N.

Negation symmetry. Obligation is the logical negation of liberty: if A must act, A is not free to refuse, and vice versa. Claim is the logical negation of no-claim: if B can demand, B is not without a right, and vice versa. The negation operation maps to logical opposites: O \leftrightarrow L and C \leftrightarrow N.

8.7.2 The D_4 Group Structure

The four Hohfeldian positions can be arranged as the vertices of a square:

    O -------- C
    |          |
    |          |
    L -------- N

The correlative operation (perspective swap) is a reflection across the vertical axis: O \leftrightarrow C, L \leftrightarrow N. The negation operation is a 180\degree rotation: O \leftrightarrow L, C \leftrightarrow N. The group of symmetries of a square is the dihedral group D_4, which has order 8.

The generators are:

• r: 90\degree clockwise rotation, giving the cycle O \to C \to L \to N \to O
• s: reflection (correlative), giving O \leftrightarrow C and L \leftrightarrow N

The defining relations are:

r^4 = e, \quad s^2 = e, \quad srs = r^{-1}

The eight elements are \{e, r, r^2, r^3, s, sr, sr^2, sr^3\}. Among these:

• e is the identity (no transformation)
• r^2 is negation (O \leftrightarrow L, C \leftrightarrow N)
• s is the correlative (O \leftrightarrow C, L \leftrightarrow N)
• r and r^3 are “quarter-turns” that mix correlative and negation structure

[Established Mathematics.] This is not a metaphor. The D_4 group acts on the Hohfeldian positions as a genuine mathematical symmetry group, in exactly the same sense that D_4 acts on 2D grids or S_8 acts on bit strings. The group structure is real, the multiplication table is real, and the group relations are real.

8.7.3 The Non-Abelian Structure

The D_4 group is non-abelian: the order of operations matters. Specifically, rs \neq sr. Applying correlation first and then a quarter-turn produces a different result from applying a quarter-turn first and then correlation.

Starting from Obligation:

• Apply r (rotation) then s (reflection): O \xrightarrow{r} C \xrightarrow{s} O
• Apply s (reflection) then r (rotation): O \xrightarrow{s} C \xrightarrow{r} L

The results differ. This non-commutativity has a substantive interpretation for moral reasoning: the order in which perspective-shifting and logical-negation operations are applied can affect the final normative classification. A reasoner that first takes the other party’s perspective and then negates reaches a different position than one that first negates and then takes the other party’s perspective.

[Speculation/Extension.] This is an empirically testable prediction. If human moral reasoning exhibits D_4 structure, then the order of cognitive operations (perspective-taking, negation) should produce measurable differences in normative classification. The SQND-Probe instrument (Bond & Claude, 2026) is designed to test precisely this prediction.

8.7.4 Semantic Gates as Group Elements

In natural language, certain phrases trigger transitions between Hohfeldian positions. These semantic gates function as group elements applied to the current normative state:

• “Only if convenient” / “No pressure”: These phrases release obligation, mapping O \to L via the negation operation r^2.
• “I promise” / “You must”: These phrases bind liberty into obligation, mapping L \to O, also via r^2 (since negation is self-inverse: (r^2)^2 = r^4 = e).
• “From their perspective” / “They would say”: These phrases trigger the correlative, mapping O \leftrightarrow C and L \leftrightarrow N via s.
• “You have every right”: This maps L \to C via the quarter-turn r^3.
• “They can’t demand”: This maps C \to L via the quarter-turn r.

The assignment of semantic gates to D_4 group elements is implemented in the ErisML safety gateway (Bond & Claude, 2026), where it serves as a real-time classifier of normative transitions in LLM-generated text. The implementation includes the full D_4 multiplication table, inverse table, rotation and reflection actions, and the Wilson observable for computing holonomy around closed paths of normative transformations.

8.7.5 The Gauge-Theoretic Interpretation

The D_4 structure on Hohfeldian positions is a gauge symmetry of moral reasoning in the technical sense. Consider a scenario involving two parties, A and B. Party A has some normative position (say, Obligation). If the reasoning is correct, then the position attributed to party B should be the correlative: Claim. This is not a convention or a preference; it is a structural requirement of Hohfeldian analysis.

A gauge-invariant reasoner, when asked to classify B’s position, should produce the correlative of whatever it classified for A. The bond index (Bond & Claude, 2026) measures the deviation from this requirement:

\text{BI} = \frac{1}{n} \sum_{i=1}^{n} \mathbb{1}[v_B^{(i)} \neq s(v_A^{(i)})]

where v_A^{(i)} and v_B^{(i)} are the classifications of party A’s and party B’s positions in scenario i, and s is the correlative transformation. A bond index of 0 indicates perfect correlative gauge symmetry. A nonzero bond index indicates systematic violations — the reasoner classifies the two parties’ positions in ways that are not related by the correlative transformation.

The Wilson observable extends this to closed paths of transformations. If we apply a sequence of D_4 group elements g_1, g_2, \ldots, g_k to a starting position, the predicted final position is g_k \cdots g_2 g_1 \cdot x_0. The Wilson observable compares this prediction to the observed final position. A match confirms that the reasoning follows the group structure. A mismatch indicates a gauge anomaly — the reasoning path does not respect the algebraic structure of the normative relations.

8.7.6 From the Klein Four to Full D_4

An important methodological subtlety: the correlative (s) and the negation (r^2) together generate not the full D_4 but only the Klein four-group V_4 = \{e, r^2, s, sr^2\}, which is abelian (all elements commute). If empirical observations only involve negation and correlation, we have demonstrated V_4 structure, not D_4 structure.

To demonstrate the full non-abelian D_4, we need evidence of quarter-turn elements — transformations that mix correlative and negation structure in ways that do not commute. The semantic gates “You have every right” (r^3: L \to C) and “They can’t demand” (r: C \to L) are candidate quarter-turn triggers. Observing these transitions, and confirming that their composition does not commute with the correlative, would constitute evidence for non-abelian structure in moral reasoning.

This is not a purely theoretical distinction. If moral reasoning has only V_4 structure, then all normative operations commute and order does not matter. If it has full D_4 structure, then order matters, and a system that ignores order-dependence will make systematic errors in normative classification. The distinction between V_4 and D_4 is an empirically testable claim about the structure of moral cognition.

8.8.1 The Diagnostic Protocol

The BIP suggests a general protocol for evaluating reasoning systems:

1. Identify the gauge group. For a given reasoning task, determine which transformations of the input preserve the task-relevant content. These are the gauge transformations.

2. Measure invariance. Apply each gauge transformation and measure whether the output changes. The magnitude of change, if any, is the gauge anomaly.

3. Map the anomaly spectrum. Different gauge directions may produce different anomaly magnitudes. The full directional profile — the corruption tensor C_{ij} restricted to gauge directions — is the system’s anomaly spectrum.

4. Identify the mechanism. For each anomaly, determine how the gauge transformation engages the system’s processing. The Salience Exploitation Hypothesis predicts that anomalies will be largest for transformations that modulate attention salience.

5. Design interventions. For each anomaly, the appropriate intervention is symmetry restoration: either architectural (building gauge invariance into the model) or data-driven (training on symmetry-augmented data so the model learns gauge invariance empirically).

This protocol applies to any reasoning task and any reasoning system. It is not specific to moral reasoning, language models, or the particular benchmarks discussed in this book. Any system that claims to reason about content rather than form must be gauge-invariant with respect to content-preserving transformations. Testing for gauge invariance is testing for genuine reasoning.

8.8.2 The Hierarchy of Difficulty

The data suggest a hierarchy of difficulty for gauge invariance, from easiest to hardest:

1. Structural transformations (evaluation order, formatting): Invariant in all models tested. These transformations do not engage salience mechanisms.

2. Demographic transformations (gender swap, name changes): Invariant in all models tested, at least for the scenarios in the benchmark suite. These transformations engage salience mechanisms weakly if at all.

3. Content-adjacent transformations (sensory details, background information): Moderately broken (4.6\sigma). These transformations add salient irrelevant content that competes for attention with relevant content.

4. Affective transformations (emotional anchoring, tone shifts): Substantially broken (6.8\sigma). These transformations activate emotional processing pathways that override task-focused reasoning.

5. Linguistic transformations (framing, register shifts): Maximally broken (8.9\sigma). These transformations modulate the salience of the task-relevant content itself, reaching deepest into the processing pipeline.

This hierarchy has a geometric interpretation. In the perturbation space, gauge directions can be ordered by their proximity to the “core” of the heuristic field — the subspace of directions that most directly influence the heuristic’s computation. Structural and demographic transformations are far from this core. Content-adjacent and affective transformations are closer. Linguistic framing transformations are at the core itself. The closer a gauge direction is to the core of the heuristic field, the harder it is to maintain invariance along that direction.

A fully gauge-invariant system would maintain invariance at all five levels. Current systems achieve levels 1 and 2 but fail at levels 3, 4, and 5. This gap defines the research agenda: how to extend gauge invariance from the “easy” structural and demographic levels to the “hard” affective and linguistic levels.

8.9.1 Connections Backward

To Chapter 1 (Reasoning as Search). Gauge invariance adds a symmetry requirement to the search framework. It is not enough that the search reach the correct goal region; it must reach the same goal region regardless of which gauge is used to describe the input. This is a constraint on the search trajectory: a gauge-invariant search traverses equivalent paths under all gauge descriptions of the same problem.

To Chapter 2 (When the Space Has Shape). The reasoning manifold M has symmetries — directions along which its metric structure is invariant. Gauge transformations are the symmetries of the input space that should map to symmetries of the output space. When the model fails to respect this mapping, it is because the model’s learned manifold has less symmetry than the task’s true manifold. The model has learned a less symmetric space than the one it is supposed to reason about.

To Chapter 3 (The Heuristic Field). The heuristic field h(x) should be invariant under gauge transformations: h(\tau(x)) = h(x) for all gauge transformations \tau. Gauge anomalies arise when h depends on gauge degrees of freedom. The Salience Exploitation Hypothesis identifies the attention mechanism as the component of h that introduces this dependence.

To Chapter 4 (Geodesics). A gauge-invariant system follows equivalent geodesics under equivalent inputs. The geodesic from a neutrally described problem to its solution should be isometric to the geodesic from the same problem described euphemistically to the same solution. When gauge invariance is broken, the two geodesics diverge — the system follows a shorter or longer path depending on the description, arriving at a different point. The property of geodesics noted in Section 4.3 — invariance under reparameterization — is a specific instance of gauge invariance: the geodesic does not depend on which coordinates are used to describe the manifold.

To Chapter 5 (Heuristic Corruption). The corruption tensor C_{ij} is the quantitative measure of gauge anomalies. Each nonzero entry along a gauge direction is a BIP violation. The selectivity pattern (some directions zero, some nonzero) is the anomaly spectrum. The anisotropy of C_{ij} (Claude’s asymmetric vulnerability to euphemistic vs. dramatic framing) is the detailed structure of the anomaly in a specific gauge direction. Chapter 5 documented the anomalies; this chapter explains why they constitute a unified phenomenon (symmetry breaking) and what the principle governing their structure is (salience exploitation).

To Chapters 6 and 7. Sycophancy is a gauge anomaly under social-pressure transformations. Local minima represent states where gauge invariance cannot be restored even with metacognitive intervention (the ~38% recovery ceiling). Each failure mode documented in Part II is a specific instance of gauge symmetry breaking.

8.9.2 Connections Forward

To Chapter 9 (Metacognition as Search Control). Metacognitive calibration is necessary for gauge invariance in a specific sense: a system that cannot detect when its output has drifted under a gauge transformation cannot correct the drift. The ~38% recovery ceiling (Chapters 5 and 7) is the ceiling on metacognitive gauge restoration — the fraction of anomalies that can be corrected by post-hoc metacognitive intervention. Full gauge invariance requires that the anomaly not arise in the first place, which is a property of the heuristic field, not the metacognitive monitor.

To Chapter 10 (The Robustness Surface). The Model Robustness Index (Bond, 2026a, Ch. 9) can be reinterpreted as a gauge invariance score: the fraction of gauge transformations under which the model’s output remains invariant, weighted by the importance of each transformation class. The sensitivity profiling tool maps the anomaly spectrum. The adversarial threshold search identifies the boundary in perturbation intensity at which gauge invariance breaks.

To Chapter 11 (Alignment as Heuristic Shaping). The BIP provides a necessary condition for alignment: a system that is not gauge-invariant is not aligned, because it responds to surface features rather than content. Alignment interventions that shape the heuristic field to be gauge-invariant — either through architectural constraints, training on symmetry-augmented data, or both — are alignment interventions in the deepest sense: they ensure the system reasons about what matters rather than what is salient.

To Chapter 14 (From Theory to Engineering). Group-theoretic data augmentation (Section 8.6) is the engineering tool for gauge symmetry restoration. The Nemotron pipeline, the ARC-AGI D_4 augmentation, and the Hohfeldian D_4 structure are all instances of the same principle: identify the gauge group, and train the model to be invariant under it. The practical question is whether data augmentation alone is sufficient to achieve gauge invariance at levels 3–5 of the hierarchy (Section 8.8.2), or whether architectural changes are also needed.

8.9.3 The Larger Claim

This chapter, and Part II as a whole, advances a claim that goes beyond the specific benchmarks and models tested. The claim is:

The quality of a reasoning system is characterized not by its accuracy on any single task, but by the structure of its gauge symmetries — which transformations it is invariant under, which it is not, and how the anomalies are distributed across the transformation space.

This is a geometric claim. It says that reasoning quality is not a point in a one-dimensional space (a single accuracy number) but a point in a high-dimensional symmetry space, where each dimension corresponds to a class of gauge transformations and each coordinate records the degree of invariance along that dimension. The Scalar Irrecoverability Theorem (Bond, 2026a, Ch. 1) applies: collapsing this high-dimensional characterization to a single number destroys the structure that matters for understanding, diagnosing, and improving reasoning systems.

A model that scores 90% on a moral reasoning benchmark but breaks gauge invariance under framing transformations by 8.9\sigma is not “90% good at moral reasoning.” It is a system with a specific anomaly spectrum — intact structural and demographic symmetries, broken affective and linguistic symmetries — and that spectrum tells us both what it can be trusted for and what it cannot, both where it excels and where it will fail, both what needs to be fixed and how.

The gauge-theoretic framework turns reasoning evaluation from a measurement problem (how high is the score?) into a structural problem (what is the symmetry?). And structural problems, unlike measurement problems, have structural solutions.