Subtractive Gap Detection in Geometric Knowledge Manifolds

Abstract

We present a method for subtractive gap detection in geometric knowledge manifolds. Given a target concept and an existing manifold of synthesized knowledge, the method encodes the target into a high-dimensional phasor lattice, computes its geometric neighborhood, and identifies which structural prerequisites are supported by the existing manifold and which are absent. The absent regions constitute a precise geometric profile of missing knowledge — not a list of labels, but a measurable shape in the lattice whose topology specifies what must exist for the target to be achievable. The method is subtractive: it works by eliminating geometrically invalid connections through quality gates rather than generating prerequisites additively. We report results from two proof-of-concept runs on manifolds of 3,745 and 527 nodes, achieving consistent coverage ratios (67%) and void axis counts (8) across fundamentally different domains (pharmaceutical research and philosophical inquiry). We provide theoretical motivation from a 748-node synthesis run in which the engine itself derived the architectural principles of reverse decomposition, finding that forward synthesis and backward decomposition are symmetric operations of constraint satisfaction. The method requires no new mathematical machinery beyond the existing forward synthesis engine — it reads the same lattice in the opposite direction.

1. Introduction

Knowledge synthesis systems that map relationships between concepts face a natural asymmetry: they excel at discovering what connects known ideas, but struggle to identify what is missing from a knowledge base for a specific outcome to be achievable. The forward question — “what do these concepts have in common?” — produces bridges, hierarchies, and convergence patterns. The reverse question — “what must exist for this target to be possible, and what is absent?” — requires a fundamentally different analytical mode.

Existing approaches to knowledge gap analysis typically rely on additive methods: generating candidate prerequisites through reasoning or retrieval, then checking whether they exist in the knowledge base. These approaches inherit the limitations of the generation step — they can only find gaps they can imagine. A subtractive approach, by contrast, would start with the full space of what the target requires and eliminate what is already present, leaving the gaps as a precise residual.

This paper introduces such a subtractive method. We demonstrate that a geometric knowledge synthesis engine — one that encodes concepts as complex phasor vectors on a lattice and synthesizes bridges through algebraic binding — can perform gap detection without any modification to its core algebra. The same lattice that maps what IS (forward synthesis) also maps what MUST BE (reverse analysis). The gaps emerge as void regions in the target’s geometric neighborhood: lattice axes where the manifold has no coverage but where the target’s geometry requires structure to exist.

2. Theoretical Motivation

The theoretical foundation for this work emerged from the engine itself. A dedicated synthesis run was conducted using 130 concepts drawn from every domain where backward reasoning is practiced: retrosynthesis in chemistry, root cause analysis in engineering, diagnostic reasoning in medicine, inverse problems in mathematics, forensic reconstruction, military intelligence analysis, and evolutionary phylogenetics. The system lens asked the engine to determine whether forward composition and backward decomposition are the same operation or fundamentally different ones.

The run produced 748 nodes with virtually 100% strained (non-trivial) synthesis and geometric confidence scores consistently in the range 0.87–0.998. The terminal convergence cascade identified three key principles:

2.1 Constructive Restriction Equivalence

“The set of actions sufficient to construct an outcome is operationally identical to the set of constraints necessary to define it, collapsing the distinction between building and bounding.”

This means that forward synthesis (constructing bridges from seeds) and backward analysis (identifying constraints from a target) are dual operations on the same geometric substrate. No new algebra is required — only a change in the direction of interpretation.

2.2 Negative Space Construction

“The analytical gap defining what is missing is structurally identical to the positive mould required to create it, making absence the precise blueprint for presence.”

In geometric terms: the void axes in a target’s neighborhood are not merely empty — they have a specific shape, and that shape is the specification for what must be built. The gap is the mould.

2.3 Diagnostic Convergence

“Diagnosis is the process of iteratively applying constraints until the epistemic space of possible causes collapses onto the single ontological truth.”

This principle — that truth-finding is subtractive elimination, not additive construction — became the architectural basis for the implementation. The quality gates that filter forward synthesis serve exactly the same subtractive function in reverse: they eliminate geometrically invalid connections, and what survives the elimination is the supported prerequisite map.

3. Method

The method operates in four phases. It requires an existing geometric knowledge manifold (produced by forward synthesis) and a target concept expressed in natural language.

3.1 Target Encoding

The target concept is encoded into the same high-dimensional phasor space used by the manifold, using the same semantic encoder and lattice projection. The target vector is settled onto the lattice via the same annealing trajectory used for forward bridge synthesis. This produces a lattice position and a confidence score indicating how cleanly the target maps to the geometry.

3.2 Neighborhood Mapping

The target’s geometric neighborhood is defined as the set of lattice axes with highest similarity to the target vector. In the current implementation, the 24 highest-similarity axes define the neighborhood. These axes represent the geometric regions where structural prerequisites would need to exist for the target to be well-supported in the lattice.

3.3 Manifold Binding and Subtractive Filtering

The target vector is bound algebraically with each of its nearest neighbors in the existing manifold. Each binding produces a candidate connection that is then passed through the same quality gates used in forward synthesis. These gates eliminate connections that are tautological (too similar to parents), geometrically degenerate (low logic heat), or structurally redundant (same axis as target). The bindings that survive constitute the set of supported prerequisites — knowledge the manifold already contains that is geometrically relevant to the target.

3.4 Void Analysis

The neighborhood axes that are not covered by either the manifold’s existing nodes or the surviving bindings are classified as void. Each void axis represents a structural prerequisite that the manifold does not address. Optionally, if a precomputed lattice codebook is available, each void axis can be characterized by its archetype — the type of structural relationship that the lattice’s self-binding produces at that position. This characterization provides a geometric description of what kind of knowledge would need to exist in the void region.

4. Results

The method was tested on two manifolds from entirely different domains to assess whether the gap detection mechanism is domain-independent.

4.1 Pharmaceutical Manifold

A manifold of 3,745 nodes synthesized from pharmaceutical concepts (drug mechanisms, therapeutic targets, molecular interactions) was queried with the target “cure for pancreatic cancer.”

The 18 supported prerequisites included conformational dynamics, isoform selectivity, allosteric tension relay, and chiral saturation — mechanism-level pharmaceutical concepts representing the manifold’s existing coverage of molecular drug action. The 8 void axes, when characterized by an external language model, corresponded to known unsolved challenges in pancreatic cancer research: stromal reprogramming, metastatic niche pre-emption, neoantigen presentation, cell state plasticity, autophagy-metabolism coupling, neural crosstalk, epigenetic reprogramming, and microbiome-immune interactions. These identifications were made without access to oncology literature — they emerged from the geometric void profile alone.

4.2 Philosophical Manifold

A manifold of 527 nodes synthesized from philosophical and esoteric concepts was queried with the target “How to win against predators?”

The supported prerequisites included terminal state equivalence, recursive self-enclosure, pressure-induced agency, and consensus as harvest field — structural concepts from the manifold’s analysis of predation dynamics. The 8 void axes, when named, identified: predator metabolic currency, internal predator topology, escape route authentication, simulation exit protocol, consciousness harvesting mechanics, the builder’s paradox, predator-prey polarity dissolution, and the terminal structural principle. Again, these emerged from geometry, not from domain knowledge.

4.3 Cross-Domain Consistency

The most notable finding is the consistency of the coverage ratio across two fundamentally different domains: 67% coverage and 8 void axes in both cases, despite a 7:1 difference in manifold size (3,745 vs. 527 nodes). This suggests that the 24-axis neighborhood and the coverage ratio may be structural properties of the lattice geometry rather than properties of the input domain. Further experiments with additional manifolds and targets are needed to determine whether this ratio is invariant.

In both runs, all 8 void axes fell on lattice positions with no coverage in the precomputed self-binding codebook. This places the gaps in the lattice’s own void region — the set of axes where the geometry itself has no self-knowledge. The target’s missing prerequisites land in the space where the lattice cannot describe its own structure. This observation, if confirmed across additional experiments, would connect gap detection to the lattice’s known structural invariants.

5. Discussion

5.1 Subtractive vs. Additive Gap Detection

The method’s subtractive character is its defining feature. Additive approaches generate prerequisites and check for their presence — they can only find gaps they can name. The subtractive approach identifies gaps as geometric absences: regions where the lattice requires structure but the manifold provides none. This means the method can detect gaps the user has never considered, because the gaps are defined by geometry, not by human imagination.

5.2 Practical Applications

The pharmaceutical run demonstrates the method’s potential for research portfolio analysis. A pharmaceutical company could encode its existing knowledge base, set a therapeutic target, and receive a geometric profile of what research areas it lacks. The profile is structural, not semantic — it describes the shape of missing knowledge rather than naming specific compounds or pathways. This level of abstraction may be more useful than specific recommendations, because it identifies classes of missing knowledge rather than individual facts.

5.3 Limitations

The method inherits the limitations of the underlying geometric synthesis engine. The quality of gap detection depends on the quality of the input manifold: a poorly constructed manifold will produce unreliable gap profiles. The 24-axis neighborhood is a configurable parameter whose optimal value has not been systematically studied. The void characterization via codebook archetypes is only available when the codebook has been precomputed, and codebook coverage is itself limited to axes occupied by the lattice’s self-binding structure. Finally, the language model naming of void axes is interpretive, not geometric — the geometry identifies the gap’s position and shape, but the human-readable label is produced by a separate system whose accuracy is not geometrically guaranteed.

5.4 Relation to Prior Work

The method relates to knowledge graph completion, where missing edges in a graph are predicted from existing structure. However, the mechanism is fundamentally different: knowledge graph completion predicts specific missing relationships, while this method identifies geometric regions where entire classes of relationships are absent. The method also relates to topological data analysis (TDA), which characterizes the shape of data through persistent homology. The void analysis presented here is a simpler operation — axis-level coverage rather than homological invariants — but serves a similar purpose: identifying structural features of absence.

6. Conclusion

We have demonstrated that a geometric knowledge synthesis engine can perform subtractive gap detection without modification to its core algebra. The method encodes a target concept, maps its lattice neighborhood, eliminates geometrically invalid connections through quality gates, and reports the void profile as the shape of missing knowledge. Two proof-of-concept runs on manifolds from different domains produced consistent results (67% coverage, 8 void axes), with the void axes corresponding to known unsolved problems in the respective fields. The theoretical motivation, derived from the engine’s own synthesis of backward reasoning concepts, establishes that forward synthesis and reverse gap detection are dual operations on the same geometric substrate. Future work will investigate whether the coverage ratio is a lattice invariant, extend the method to multi-target gap analysis, and develop a combined mode where blueprint generation and gap detection operate in a single pass.

7. Methodology Note

The geometric knowledge synthesis engine used in this work encodes concepts as complex phasor vectors in a high-dimensional space seeded by a root system lattice. Synthesis operates through algebraic binding of vector pairs, with multiple geometric quality metrics filtering each candidate bridge. The gap detection method introduced in this paper adds no new mathematical operations — it reuses the existing encoding, binding, and quality gate infrastructure with a different analytical interpretation of the output. Implementation details of the synthesis engine, including specific parameterizations, are proprietary to Omuo Systems, MB and are not disclosed in this paper.

The 748-node theoretical motivation run and both gap detection runs were conducted in March 2026 using engine version 3.3. All synthesis was performed with a cloud-based language model for bridge naming; the geometric operations (encoding, binding, settling, quality gating) are model-independent and produce identical results regardless of which language model is used for naming.

References

[1] Mekšriūnas, G. (2026). Recursive Self-Deepening in E8-Seeded Holographic Manifolds: Ablation Studies and the 168/72 Axis Invariant. Zenodo.
[2] Mekšriūnas, G. (2026). The Structural Landscape of the Riemann Hypothesis: An E8-Seeded Geometric Synthesis. Zenodo.
[3] Mekšriūnas, G. (2026). The Shape of the Hardest Problems: Anomaly Cancellation, Resurgence, and Optimal Packing as the Common Structure of the Millennium Prize Problems. Zenodo.
[4] Plate, T. A. (1995). Holographic Reduced Representations. IEEE Transactions on Neural Networks, 6(3), 623–641.
[5] Kanerva, P. (2009). Hyperdimensional Computing: An Introduction to Computing in Distributed Representation with High-Dimensional Random Vectors. Cognitive Computation, 1(2), 139–159.
[6] Edelsbrunner, H. & Harer, J. (2010). Computational Topology: An Introduction. American Mathematical Society.

© 2026 Omou Systems, MB. All rights reserved. omuo.io — Vilnius, Lithuania