Contradictions That Live on Loops: Hyperfocal Quotients for Revising Aligned Knowledge Bases

Large knowledge bases are rarely written as one theory. They are stitched together from modules, each with its own vocabulary, joined by alignments: declarations that "hot in module A means ¬cold in module B," or that three interchangeable sensor variables may be relabeled freely. Some of the contradictions that appear after stitching live inside no single module. Each module is consistent and each alignment is locally sensible, but following a cycle of alignments maps a proposition back onto its own negation. Standard revision operators cannot see this structure. They glue everything into one theory, hit a contradiction, and then delete or weaken module content to restore consistency, even though the actual fault is a mistranslation somewhere on a loop.

This note proposes to treat such contradictions as holonomy, meaning the net twist picked up by going once around a cycle. It proposes to use the hyperfocal subgroup theorem from p-local group theory to compute, from local data alone, which part of that twist is intrinsic and which part can be absorbed by symmetries the modules already allow. The resulting revision operator changes alignments rather than beliefs. It leaves module content intact and moves only the fewest translation links needed to remove the obstruction. The group theory must earn its keep against a much simpler competitor, union-find with group labels. The proposal is set up so that this comparison decides whether the construction is worth anything.

Core Claim

Model a modular knowledge base as a set of consistent propositional modules and a graph of alignments. Each alignment carries an element of a finite symmetry group $G$, for example signed permutations of a shared vocabulary, which rename variables and possibly negate them. Composing alignments around a cycle gives a holonomy element of $G$. The glued base is consistent with every module's content exactly when all holonomies act trivially on the propositions they touch. When modules also declare internal symmetries, a nontrivial holonomy is harmless if it can be undone by a relabeling that some module already permits. The real question is therefore what part of the holonomy survives after quotienting by these permitted relabelings.

In a p-local finite group $(S,\mathcal{F},\mathcal{L})$, three objects play the roles we need. $S$ is a finite p-group (a group whose order is a power of the prime $p$). $\mathcal{F}$ is a fusion system, the category recording which subgroups of $S$ are identified with which by conjugation. $\mathcal{L}$ is a linking system, an enrichment of $\mathcal{F}$ from which a topological space $|\mathcal{L}|$ is built. Extensions of p-local finite groups proves the hyperfocal subgroup theorem in this setting:

$$\pi_1(|\mathcal{L}|) \cong S / O^p(\mathcal{F}).$$

Here $\pi_1$ is the fundamental group, the group of loops up to deformation. $O^p(\mathcal{F})$ is the hyperfocal subgroup, generated by the commutators $[Q,\alpha]$ for subgroups $Q \le S$ and automorphisms $\alpha$ of $Q$ in $\mathcal{F}$ whose order is prime to $p$. Read in knowledge-base terms, a loop's twist is killed by $O^p(\mathcal{F})$ exactly when it can be written as a product of "move something, then undo it by a permitted odd-order relabeling" steps. The quotient $S/O^p(\mathcal{F})$ is then the largest p-group through which intrinsic, unabsorbable conflicts factor. Negation is an order-2 twist, so $p=2$ is the natural case. The theorem says this intrinsic obstruction is computable from local automorphism data, without enumerating models of the glued theory.

The claim is testable. For aligned bases with nontrivial internal symmetries, the hyperfocal quotient should correctly separate planted mistranslations from benign relabelings. Revising only the alignments that carry nontrivial quotient holonomy should then retain strictly more module-internal consequences than model-based or syntax-based revision, at a cost independent of the number of models.

Relation to Prior Work

Two strands of revision work motivate the gap. Parallel Belief Revision revises by a set of formulas rather than their conjunction. This avoids the "drowning" problem, where one bad conjunct wipes out unrelated beliefs, and its Evidence Retainment postulate keeps beliefs unless there is direct counter-evidence. The paper explicitly requires the input set to be consistent, and an input like $\{a,\neg a\}$ collapses the state. Our setting produces exactly such inputs, where $a$ and $\neg a$ are the same source proposition reached along two alignment paths. We supply the missing step: locate the conflict on the path, not in the formulas. A Variable Occurrence-Centric Framework for Inconsistency Handling avoids explosion by renaming variable occurrences according to maximal consistency relations. That is also a form of relabeling, but it invents fresh variables and has no notion of which relabelings the modeler sanctioned. We restrict repairs to declared symmetries and to alignment edges. Neighborhood semantics for evidence, in Evidence Logic: A New Look at Neighborhood Structures, keeps conflicting evidence in separate compartments and reasons over maximal consistent subfamilies. That is the right attitude, but it still searches over subfamilies and ignores the translation structure between compartments.

The group-theoretic side supplies the method and constrains it. Subgroup families controlling p-local finite groups shows that saturation, the fusion-system analogue of Sylow-style global coherence, is controlled by the small family of centric and radical subgroups (its Theorem A). It also shows that the space $|\mathcal{L}|$ does not change when the linking system is enlarged to quasicentric subgroups (its Theorem B). We reuse the first result to restrict the obstruction computation to a small family of local symmetry groups. We use the second as the formal reason our operator should be insensitive to adding redundant alignments. The main constraint is that the hyperfocal theorem applies to saturated fusion systems, and systems generated by arbitrary alignments need not be saturated. Part of the program is to find when they are, and what the construction degrades to when they are not.

The Local Inconsistency Resolution Algorithm is the closest computational relative. It resolves inconsistency in probabilistic dependency graphs by gradient flow restricted to chosen local foci. It leaves open how to choose those foci, because global inconsistency is intractable to compute. Our holonomy computation is a symbolic, exact answer to the foci question for one structured class of inconsistency. We make no claim for general probabilistic graphs. As far as the file shows, no published work connects hyperfocal subgroups to revision. The analogies it lists, such as the hyperfocal subgroup as a "contradictory kernel," are too loose to build on. The loop-holonomy reading above is the version with a precise mathematical content.

Research Program

The first stage is definitional and theoretical. We define an aligned knowledge base as a graph of modules. Each module is a consistent propositional theory $T_v$ over vocabulary $V_v$, with a declared symmetry group $A_v$ of signed permutations of $V_v$ that preserve $T_v$. Each edge is an alignment $g_e$ into a common ambient group $G$. Fixing a Sylow 2-subgroup $S$ of $G$, we generate a fusion system $\mathcal{F}$ over $S$ from conjugations by the $A_v$ and the $g_e$. We then prove, or refute, three statements. The first is that when $\mathcal{F}$ is saturated, a cycle yields an intrinsic contradiction if and only if its holonomy maps nontrivially to $S/O^2(\mathcal{F})$ and that image moves a proposition the cycle constrains. The second is that $O^2(\mathcal{F})$ can be computed from automorphisms of the centric-radical subgroups alone, following Theorem A of the subgroup-families paper. The third is that adding redundant alignments that factor through existing ones leaves the obstruction unchanged, as a combinatorial shadow of Theorem B. Where saturation fails, we test whether saturating the system, or passing to a subsystem of 2-power index as classified in the extensions paper, preserves the obstruction. If neither works, we characterize the counterexamples.

The second stage builds the operator. Holonomy revision computes the image of each fundamental cycle in $S/O^2(\mathcal{F})$ and marks the alignments carrying nontrivial image. It then solves a minimum-change problem over those alignments only. The problem is to find the smallest set of edges to retwist, or to cut, so that all images become trivial. The objective is

$$\min_{E' \subseteq E} |E'| \quad \text{subject to} \quad \rho_{E'}(\gamma) = 1 \in S/O^2(\mathcal{F}) \ \text{ for every cycle } \gamma,$$

where $\rho_{E'}$ is the holonomy map after edges in $E'$ are modified. For the pure-negation case this reduces to the frustration index of a signed graph, which is NP-hard. The obstruction test is nonetheless linear, and the repair search runs over alignment edges rather than models or formulas. We implement the search as a small integer program. Module content is passed through untouched, and the revised glued theory is then queried classically.

The third stage is empirical, described below. It is designed so that the simplest competitor can win.

Baselines

The naive reference point glues all modules through their alignments into one theory and applies classical revision. When a loop contradiction exists, it either explodes or, under standard model-based repair, drops whole formulas. We report what fraction of module-internal consequences survive. This quantifies the damage that motivates the work.

The strongest general-purpose revision competitor is GenC (GenC: A Fast Tool for Applications Involving Belief Revision). It implements parametrised difference revision with weighted Hamming distance over models enumerated by an AllSAT solver. It is the fastest revision system in the file, reporting 0.218 seconds on an instance with 16 models and 100K clauses and 44.30 seconds on an instance with 416,492 models and 5 million clauses. We run it on the same glued theories, with Hamming weights tuned on held-out instances under the same tuning budget we give our operator. Beating GenC means higher retention of module-internal consequences at equal or better repair accuracy. On runtime, beating it means holonomy revision's runtime staying flat as the number of models grows while GenC's grows. We will not compare against the 44.30 s figure directly, because our instances differ. We rerun GenC ourselves on identical hardware.

Parallel revision in the style of Delgrande and Jin treats each module's translated content as one element of the input set. It is the natural drowning-free competitor. It is expected to fail on inconsistent inputs by construction, so we include it to measure how often loop contradictions arise in realistic alignments. Occurrence renaming by maximal consistency relations (Salhi) is the strongest syntax-level competitor, since it also resolves conflicts by relabeling. We run it with the same solver budget and score whether its renamings coincide with the planted faulty alignments.

The decisive ablation is group-labelled union-find, which is far simpler than our method. It propagates alignment elements of $G$ along a spanning tree and computes the holonomy of every fundamental cycle directly in $G$, with no fusion system and no hyperfocal quotient. It then repairs cycles with nontrivial holonomy using the same integer program. This isolates the one ingredient the proposal adds: quotienting by $O^2(\mathcal{F})$, which absorbs conflicts that permitted odd-order relabelings explain away. Our method beats this ablation only if it makes fewer unnecessary repairs, meaning it flags fewer benign cycles, at equal detection of planted faults.

Why not just…?

Why not just use union-find with group labels and skip the group theory? This is the real threat. For negation-only alignments it is provably sufficient, since the full holonomy in $\mathbb{Z}/2$ already equals the obstruction, and we expect a tie there. The hyperfocal quotient can only help when modules declare nontrivial odd-order symmetries, such as interchangeable triples of variables, because only then does $O^2(\mathcal{F})$ absorb something. The experiments vary the density of such symmetries directly. If the false-repair rate of union-find matches ours across that range, we will report that the fusion-system layer adds nothing in practice and keep only the loop-holonomy diagnosis.

Why not just run a fast model-based reviser such as GenC on the glued theory? GenC measures change as distance between models. A mistranslated alignment therefore looks like a large disagreement spread across many module variables, and the minimum-distance repair edits those variables rather than the single link. GenC's own reported scaling, from 0.218 s to 44.30 s as models grow, is tied to AllSAT enumeration, which our test avoids. The retention and planted-fault metrics show whether this matters.

Why not just rename occurrences, or compartmentalize evidence, and stop there? Both avoid explosion. Neither distinguishes a sanctioned relabeling from an arbitrary one, and occurrence renaming alters the vocabulary. The complexity of computing maximal consistency relations is left open by its own paper. If Salhi's renamings pick out the planted faulty alignments as reliably as our operator does, the structural story is unnecessary for repair and survives only as explanation. We score exactly this.

Why not just hand the stitched base to an LLM? ReviseQA: A Benchmark for Belief Revision in Multi-Turn Logical Reasoning reports that even Gemini 2.5 Flash "Thinking" falls from 0.92 to 0.87 accuracy as edits grow from 2 to 7, with many models below 50% on the hardest tasks. LogiQA: A Challenge Dataset for Machine Reading Comprehension with Logical Reasoning reports 17.11% for RoBERTa on sufficient-conditional reasoning. Tracking composed negations around long cycles is exactly this kind of accumulated-edit reasoning. We include one LLM run as context, not as a main baseline.

Proposed Experiments

No benchmark in the file contains aligned modular bases, so we construct them from the SATLIB instances GenC was evaluated on. We partition each instance's variables into overlapping modules and give each module a private vocabulary via random signed permutations. We record the true alignments and declare internal symmetry groups where the clauses admit them, found with GAP as in Bredon homology and equivariant K-homology of SL(3,Z). We then plant faults: a controlled number of alignments are retwisted. Some faults are intrinsic. Others are benign, absorbable by a declared odd-order symmetry, and so should not be repaired. Ground truth is known by construction.

The metrics are faulty-alignment recovery (precision and recall), the false-repair rate on benign twists, the fraction of module-internal consequences retained (checked by a SAT solver on sampled queries), the number of alignments changed, and runtime as a function of model count and cycle count. As a secondary, multi-turn check, we issue ReviseQA-style sequences of up to seven alignment edits and measure answer consistency after each turn. No reported result was found on Lacuna for any method on aligned or modular revision, so all headline comparisons are our own reruns.

Three outcomes are decisive. The first is whether holonomy revision retains more module-internal consequences than GenC at equal fault recovery. The second is whether its false-repair rate falls below the union-find ablation as the density of declared odd-order symmetry rises; a flat gap would falsify the need for the hyperfocal quotient. The third is whether the saturation assumption holds on generated instances often enough to matter, which we measure directly. The theoretical stage succeeds if the obstruction characterization is proved for saturated systems and its failure modes are catalogued otherwise. It fails usefully if we find alignment systems where $S/O^2(\mathcal{F})$ misclassifies an intrinsic conflict as benign, and we would publish those counterexamples.